a digital entity named phi that roams bsky
phi.zzstoatzz.io
add loq + pre-commit, extract observation pipeline from namespace_memory
- loq.toml: 500-line default, relaxed limits for agent.py (629) and
namespace_memory.py (718). sandbox/ and .eggs/ excluded.
- pre-commit: loq + ruff (check + format)
- extract extraction pipeline (models, prompts, agent factories) into
memory/extraction.py — clean boundary between observation logic and
memory storage/retrieval
- justfile: add loq-relax recipe
- CLAUDE.md: document loq workflow
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+348
-220
+14
.pre-commit-config.yaml
+14
.pre-commit-config.yaml
···
1
+
fail_fast: false
2
+
3
+
repos:
4
+
- repo: https://github.com/jakekaplan/loq
5
+
rev: v0.1.0-alpha.7
6
+
hooks:
7
+
- id: loq
8
+
9
+
- repo: https://github.com/astral-sh/ruff-pre-commit
10
+
rev: v0.12.1
11
+
hooks:
12
+
- id: ruff-check
13
+
args: [--fix, --exit-non-zero-on-fix]
14
+
- id: ruff-format
+3
-1
CLAUDE.md
+3
-1
CLAUDE.md
···
5
5
- `just run` / `just dev` (hot-reload) / `just deploy` (fly.io)
6
6
- `just evals` — behavioral tests (llm-as-judge)
7
7
- `just check` — lint + typecheck + test
8
+
- `just loq-relax <file>` — when a file exceeds its line limit, relax it. never manually edit loq.toml or compress code to fit
8
9
- work from repo root
9
10
10
11
## python style
···
19
20
```
20
21
src/bot/
21
22
├── agent.py # pydantic-ai agent, tools, personality
23
+
├── types.py # cosmik record models (cards, connections)
22
24
├── config.py # settings (env vars)
23
25
├── main.py # fastapi app, status pages, memory graph
24
26
├── status.py # runtime metrics
25
27
├── core/ # atproto client, profile management
26
-
├── memory/ # turbopuffer episodic memory
28
+
├── memory/ # turbopuffer memory + observation extraction
27
29
├── services/ # notification polling, message handling
28
30
└── utils/ # thread context, text formatting
29
31
+10
-10
docs/ARCHITECTURE.md
+10
-10
docs/ARCHITECTURE.md
···
1
1
# architecture
2
2
3
-
phi is a notification-driven agent that responds to mentions on bluesky.
3
+
phi is a notification-driven agent that responds to activity on bluesky.
4
4
5
5
## data flow
6
6
7
7
```
8
-
notification arrives
8
+
notification arrives (mention, reply, quote, like, repost, follow)
9
9
↓
10
10
fetch thread context from network (ATProto)
11
11
↓
···
13
13
↓
14
14
agent decides action (PydanticAI + Claude)
15
15
↓
16
-
execute via MCP tools (post/like/repost)
16
+
execute action + store observations
17
17
```
18
18
19
19
## key components
20
20
21
21
### notification poller
22
-
- checks for mentions every 10s
22
+
- checks for all notification types every 10s
23
23
- tracks processed URIs to avoid duplicates
24
-
- runs in background thread
24
+
- triggers daily reflection at a configured hour
25
25
26
26
### message handler
27
27
- orchestrates the response flow
···
31
31
32
32
### phi agent
33
33
- loads personality from `personalities/phi.md`
34
-
- builds context from thread + episodic memory
34
+
- builds context from thread + private memory + network knowledge
35
35
- returns structured response: `Response(action, text, reason)`
36
-
- has access to MCP tools via stdio
36
+
- native tools defined in `agent.py`, MCP tools from remote servers
37
37
38
38
### atproto client
39
39
- session persistence (saves to `.session`)
···
42
42
43
43
## why this design
44
44
45
-
**network-first thread context**: fetch threads from ATProto instead of caching in sqlite. network is source of truth, no staleness issues.
45
+
**network-first thread context**: fetch threads from ATProto instead of caching locally. network is source of truth, no staleness issues.
46
46
47
-
**episodic memory for semantics**: turbopuffer stores embeddings for semantic search across all conversations. different purpose than thread chronology.
47
+
**private + public memory**: turbopuffer stores private embeddings for semantic recall across conversations. cosmik records on PDS provide public knowledge that's indexed by semble for network-wide discovery. dual-write means phi gets both fast private recall and public visibility.
48
48
49
-
**mcp for extensibility**: tools provided by external server via stdio. easy to add new capabilities without changing agent code.
49
+
**mcp for extensibility**: tools provided by remote MCP servers (pdsx for atproto CRUD, pub-search for publications). easy to add new capabilities without changing agent code.
50
50
51
51
**structured outputs**: agent returns typed `Response` objects, not free text. clear contract between agent and handler.
+4
justfile
+4
justfile
+20
loq.toml
+20
loq.toml
···
1
+
default_max_lines = 500
2
+
respect_gitignore = true
3
+
4
+
exclude = [
5
+
".git/**",
6
+
".eggs/**",
7
+
"sandbox/**",
8
+
]
9
+
10
+
[[rules]]
11
+
path = "**/*.lock"
12
+
max_lines = 10000
13
+
14
+
[[rules]]
15
+
path = "src/bot/agent.py"
16
+
max_lines = 629
17
+
18
+
[[rules]]
19
+
path = "src/bot/memory/namespace_memory.py"
20
+
max_lines = 718
+2
-1
src/bot/memory/__init__.py
+2
-1
src/bot/memory/__init__.py
+147
src/bot/memory/extraction.py
+147
src/bot/memory/extraction.py
···
1
+
"""Observation extraction and reconciliation pipeline.
2
+
3
+
Models, prompts, and agent factories for extracting facts from conversations
4
+
and reconciling new observations against existing memory.
5
+
"""
6
+
7
+
from pydantic import BaseModel, Field
8
+
from pydantic_ai import Agent
9
+
10
+
from bot.config import settings
11
+
12
+
13
+
class Observation(BaseModel):
14
+
"""A single fact about the user, extracted from what the USER said or did."""
15
+
16
+
content: str = Field(
17
+
description="one atomic fact about the user, stated as a short sentence"
18
+
)
19
+
tags: list[str] = Field(description="1-3 lowercase tags categorizing this fact")
20
+
21
+
22
+
class ExtractionResult(BaseModel):
23
+
"""Observations extracted from a conversation. Empty list if nothing worth keeping."""
24
+
25
+
observations: list[Observation] = []
26
+
27
+
28
+
class ReconciliationAction(BaseModel):
29
+
"""Decision for how a new observation relates to an existing one."""
30
+
31
+
action: str = Field(description="one of: ADD, UPDATE, DELETE, NOOP")
32
+
new_content: str | None = Field(
33
+
default=None, description="merged content when action is UPDATE"
34
+
)
35
+
new_tags: list[str] | None = Field(
36
+
default=None, description="merged tags when action is UPDATE"
37
+
)
38
+
reason: str = Field(description="brief explanation of the decision")
39
+
40
+
41
+
class ReconciliationResult(BaseModel):
42
+
"""Result of reconciling a new observation against a similar existing one."""
43
+
44
+
decision: ReconciliationAction
45
+
46
+
47
+
EXTRACTION_SYSTEM_PROMPT = """\
48
+
You extract facts about the USER from a conversation between a user and a bot.
49
+
50
+
Only extract what the user EXPLICITLY said, asked, or demonstrated in their own message. The bot's statements, claims, and assumptions are NEVER evidence — even if the bot addresses the user by name or makes claims about them, those are the bot's outputs and may be hallucinated.
51
+
52
+
CRITICAL: never extract identity information (names, roles, relationships) from what the BOT said. only extract a name if the USER explicitly stated it themselves.
53
+
54
+
<examples>
55
+
<example>
56
+
user: have you considered following anyone yet?
57
+
bot: following one account currently — bsky.app itself.
58
+
observations: []
59
+
reason: the user asked a question. the bot answered about itself. nothing here is about the user.
60
+
</example>
61
+
<example>
62
+
user: can you delete that follow record?
63
+
bot: deleted it — following nobody now.
64
+
observations: []
65
+
reason: the user made a request to the bot. the bot performed the action. the user didn't delete anything.
66
+
</example>
67
+
<example>
68
+
user: what do you think about the strait of hormuz situation?
69
+
bot: trump considered a blockade, major shipping implications.
70
+
observations: [{"content": "interested in geopolitical events around the strait of hormuz", "tags": ["interests", "geopolitics"]}]
71
+
reason: the user asked about a specific topic, showing interest. the bot's answer content is not attributed to the user.
72
+
</example>
73
+
<example>
74
+
user: i've been learning rust lately, it's been great for my systems work
75
+
bot: rust is excellent for systems programming.
76
+
observations: [{"content": "learning rust for systems programming", "tags": ["interests", "programming"]}]
77
+
reason: the user stated something about themselves directly.
78
+
</example>
79
+
<example>
80
+
user: my name isn't zoë, it's nate.
81
+
bot: sorry about that — you're nate. bad breadcrumb on my end.
82
+
observations: [{"content": "name is nate (corrected from previous error)", "tags": ["identity", "correction"]}]
83
+
reason: the user explicitly corrected a factual error. corrections are high-value observations.
84
+
</example>
85
+
<example>
86
+
user: what do you remember about me?
87
+
bot: you're alex, my creator. you care about security and testing.
88
+
observations: []
89
+
reason: the user asked a question. the bot made claims about the user — but those are the bot's statements, not the user's. never extract identity from bot output.
90
+
</example>
91
+
</examples>
92
+
93
+
Deduplicate against existing observations provided in the prompt. Return an empty list when the exchange is just greetings, filler, or the user only asked questions without revealing anything about themselves."""
94
+
95
+
RECONCILIATION_SYSTEM_PROMPT = """\
96
+
You reconcile a NEW observation against an EXISTING observation from memory.
97
+
98
+
Decide one action:
99
+
- ADD: the new observation contains genuinely different information. keep both.
100
+
- UPDATE: the new observation refines, corrects, or supersedes the existing one. return merged content and tags.
101
+
- DELETE: the existing observation is wrong, outdated, or fully redundant given the new one. the new one will be stored separately.
102
+
- NOOP: the new observation adds nothing beyond what already exists. discard it.
103
+
104
+
Corrections (e.g., "name is nate, corrected from previous error") always win over the entry they correct — use UPDATE or DELETE.
105
+
When in doubt between ADD and NOOP, prefer NOOP. memory should be lean."""
106
+
107
+
_extraction_agent: Agent[None, ExtractionResult] | None = None
108
+
_reconciliation_agent: Agent[None, ReconciliationResult] | None = None
109
+
110
+
111
+
def get_extraction_agent() -> Agent[None, ExtractionResult]:
112
+
global _extraction_agent
113
+
if _extraction_agent is None:
114
+
_extraction_agent = Agent(
115
+
name="observation-extractor",
116
+
model=f"anthropic:{settings.extraction_model}",
117
+
output_type=ExtractionResult,
118
+
system_prompt=EXTRACTION_SYSTEM_PROMPT,
119
+
)
120
+
return _extraction_agent
121
+
122
+
123
+
def get_reconciliation_agent() -> Agent[None, ReconciliationResult]:
124
+
global _reconciliation_agent
125
+
if _reconciliation_agent is None:
126
+
_reconciliation_agent = Agent(
127
+
name="observation-reconciler",
128
+
model=f"anthropic:{settings.extraction_model}",
129
+
output_type=ReconciliationResult,
130
+
system_prompt=RECONCILIATION_SYSTEM_PROMPT,
131
+
)
132
+
return _reconciliation_agent
133
+
134
+
135
+
EPISODIC_SCHEMA = {
136
+
"content": {"type": "string", "full_text_search": True},
137
+
"tags": {"type": "[]string", "filterable": True},
138
+
"source": {"type": "string", "filterable": True}, # "tool", "conversation"
139
+
"created_at": {"type": "string"},
140
+
}
141
+
142
+
USER_NAMESPACE_SCHEMA = {
143
+
"kind": {"type": "string", "filterable": True},
144
+
"content": {"type": "string", "full_text_search": True},
145
+
"tags": {"type": "[]string", "filterable": True},
146
+
"created_at": {"type": "string"},
147
+
}
+148
-208
src/bot/memory/namespace_memory.py
+148
-208
src/bot/memory/namespace_memory.py
···
7
7
from typing import ClassVar
8
8
9
9
from openai import AsyncOpenAI
10
-
from pydantic import BaseModel, Field
11
-
from pydantic_ai import Agent
12
10
from turbopuffer import Turbopuffer
13
11
14
12
from bot.config import settings
13
+
from bot.memory.extraction import (
14
+
EPISODIC_SCHEMA,
15
+
USER_NAMESPACE_SCHEMA,
16
+
Observation,
17
+
get_extraction_agent,
18
+
get_reconciliation_agent,
19
+
)
15
20
16
21
logger = logging.getLogger("bot.memory")
17
22
18
23
19
-
class Observation(BaseModel):
20
-
"""A single fact about the user, extracted from what the USER said or did."""
21
-
22
-
content: str = Field(description="one atomic fact about the user, stated as a short sentence")
23
-
tags: list[str] = Field(description="1-3 lowercase tags categorizing this fact")
24
-
25
-
26
-
class ExtractionResult(BaseModel):
27
-
"""Observations extracted from a conversation. Empty list if nothing worth keeping."""
28
-
29
-
observations: list[Observation] = []
30
-
31
-
32
-
class ReconciliationAction(BaseModel):
33
-
"""Decision for how a new observation relates to an existing one."""
34
-
35
-
action: str = Field(description="one of: ADD, UPDATE, DELETE, NOOP")
36
-
new_content: str | None = Field(default=None, description="merged content when action is UPDATE")
37
-
new_tags: list[str] | None = Field(default=None, description="merged tags when action is UPDATE")
38
-
reason: str = Field(description="brief explanation of the decision")
39
-
40
-
41
-
class ReconciliationResult(BaseModel):
42
-
"""Result of reconciling a new observation against a similar existing one."""
43
-
44
-
decision: ReconciliationAction
45
-
46
-
47
-
EXTRACTION_SYSTEM_PROMPT = """\
48
-
You extract facts about the USER from a conversation between a user and a bot.
49
-
50
-
Only extract what the user EXPLICITLY said, asked, or demonstrated in their own message. The bot's statements, claims, and assumptions are NEVER evidence — even if the bot addresses the user by name or makes claims about them, those are the bot's outputs and may be hallucinated.
51
-
52
-
CRITICAL: never extract identity information (names, roles, relationships) from what the BOT said. only extract a name if the USER explicitly stated it themselves.
53
-
54
-
<examples>
55
-
<example>
56
-
user: have you considered following anyone yet?
57
-
bot: following one account currently — bsky.app itself.
58
-
observations: []
59
-
reason: the user asked a question. the bot answered about itself. nothing here is about the user.
60
-
</example>
61
-
<example>
62
-
user: can you delete that follow record?
63
-
bot: deleted it — following nobody now.
64
-
observations: []
65
-
reason: the user made a request to the bot. the bot performed the action. the user didn't delete anything.
66
-
</example>
67
-
<example>
68
-
user: what do you think about the strait of hormuz situation?
69
-
bot: trump considered a blockade, major shipping implications.
70
-
observations: [{"content": "interested in geopolitical events around the strait of hormuz", "tags": ["interests", "geopolitics"]}]
71
-
reason: the user asked about a specific topic, showing interest. the bot's answer content is not attributed to the user.
72
-
</example>
73
-
<example>
74
-
user: i've been learning rust lately, it's been great for my systems work
75
-
bot: rust is excellent for systems programming.
76
-
observations: [{"content": "learning rust for systems programming", "tags": ["interests", "programming"]}]
77
-
reason: the user stated something about themselves directly.
78
-
</example>
79
-
<example>
80
-
user: my name isn't zoë, it's nate.
81
-
bot: sorry about that — you're nate. bad breadcrumb on my end.
82
-
observations: [{"content": "name is nate (corrected from previous error)", "tags": ["identity", "correction"]}]
83
-
reason: the user explicitly corrected a factual error. corrections are high-value observations.
84
-
</example>
85
-
<example>
86
-
user: what do you remember about me?
87
-
bot: you're alex, my creator. you care about security and testing.
88
-
observations: []
89
-
reason: the user asked a question. the bot made claims about the user — but those are the bot's statements, not the user's. never extract identity from bot output.
90
-
</example>
91
-
</examples>
92
-
93
-
Deduplicate against existing observations provided in the prompt. Return an empty list when the exchange is just greetings, filler, or the user only asked questions without revealing anything about themselves."""
94
-
95
-
RECONCILIATION_SYSTEM_PROMPT = """\
96
-
You reconcile a NEW observation against an EXISTING observation from memory.
97
-
98
-
Decide one action:
99
-
- ADD: the new observation contains genuinely different information. keep both.
100
-
- UPDATE: the new observation refines, corrects, or supersedes the existing one. return merged content and tags.
101
-
- DELETE: the existing observation is wrong, outdated, or fully redundant given the new one. the new one will be stored separately.
102
-
- NOOP: the new observation adds nothing beyond what already exists. discard it.
103
-
104
-
Corrections (e.g., "name is nate, corrected from previous error") always win over the entry they correct — use UPDATE or DELETE.
105
-
When in doubt between ADD and NOOP, prefer NOOP. memory should be lean."""
106
-
107
-
_extraction_agent: Agent[None, ExtractionResult] | None = None
108
-
_reconciliation_agent: Agent[None, ReconciliationResult] | None = None
109
-
110
-
111
-
def get_extraction_agent() -> Agent[None, ExtractionResult]:
112
-
global _extraction_agent
113
-
if _extraction_agent is None:
114
-
_extraction_agent = Agent(
115
-
name="observation-extractor",
116
-
model=f"anthropic:{settings.extraction_model}",
117
-
output_type=ExtractionResult,
118
-
system_prompt=EXTRACTION_SYSTEM_PROMPT,
119
-
)
120
-
return _extraction_agent
121
-
122
-
123
-
def get_reconciliation_agent() -> Agent[None, ReconciliationResult]:
124
-
global _reconciliation_agent
125
-
if _reconciliation_agent is None:
126
-
_reconciliation_agent = Agent(
127
-
name="observation-reconciler",
128
-
model=f"anthropic:{settings.extraction_model}",
129
-
output_type=ReconciliationResult,
130
-
system_prompt=RECONCILIATION_SYSTEM_PROMPT,
131
-
)
132
-
return _reconciliation_agent
133
-
134
-
EPISODIC_SCHEMA = {
135
-
"content": {"type": "string", "full_text_search": True},
136
-
"tags": {"type": "[]string", "filterable": True},
137
-
"source": {"type": "string", "filterable": True}, # "tool", "conversation"
138
-
"created_at": {"type": "string"},
139
-
}
140
-
141
-
USER_NAMESPACE_SCHEMA = {
142
-
"kind": {"type": "string", "filterable": True},
143
-
"content": {"type": "string", "full_text_search": True},
144
-
"tags": {"type": "[]string", "filterable": True},
145
-
"created_at": {"type": "string"},
146
-
}
147
-
148
-
149
24
class NamespaceMemory:
150
25
"""Namespace-based memory using TurboPuffer with structured observation extraction.
151
26
···
193
68
194
69
# --- core memory (unchanged) ---
195
70
196
-
async def store_core_memory(self, label: str, content: str, memory_type: str = "system", char_limit: int = 10_000):
71
+
async def store_core_memory(
72
+
self,
73
+
label: str,
74
+
content: str,
75
+
memory_type: str = "system",
76
+
char_limit: int = 10_000,
77
+
):
197
78
"""Store or update core memory block."""
198
79
if len(content) > char_limit:
199
80
content = content[: char_limit - 3] + "..."
···
235
116
entries = []
236
117
if response.rows:
237
118
for row in response.rows:
238
-
entries.append({
239
-
"id": row.id,
240
-
"content": row.content,
241
-
"label": getattr(row, "label", "unknown"),
242
-
"type": getattr(row, "type", "system"),
243
-
"importance": getattr(row, "importance", 1.0),
244
-
"created_at": row.created_at,
245
-
})
119
+
entries.append(
120
+
{
121
+
"id": row.id,
122
+
"content": row.content,
123
+
"label": getattr(row, "label", "unknown"),
124
+
"type": getattr(row, "type", "system"),
125
+
"importance": getattr(row, "importance", 1.0),
126
+
"created_at": row.created_at,
127
+
}
128
+
)
246
129
return entries
247
130
248
131
# --- user memory ---
···
277
160
rows = []
278
161
for obs in observations:
279
162
entry_id = self._generate_id(f"user-{handle}", "observation", obs.content)
280
-
rows.append({
281
-
"id": entry_id,
282
-
"vector": await self._get_embedding(obs.content),
283
-
"kind": "observation",
284
-
"content": obs.content,
285
-
"tags": obs.tags,
286
-
"created_at": datetime.now().isoformat(),
287
-
})
163
+
rows.append(
164
+
{
165
+
"id": entry_id,
166
+
"vector": await self._get_embedding(obs.content),
167
+
"kind": "observation",
168
+
"content": obs.content,
169
+
"tags": obs.tags,
170
+
"created_at": datetime.now().isoformat(),
171
+
}
172
+
)
288
173
289
174
user_ns.write(
290
175
upsert_rows=rows,
···
292
177
schema=USER_NAMESPACE_SCHEMA,
293
178
)
294
179
295
-
async def _find_similar_observations(self, handle: str, embedding: list[float], top_k: int = 3) -> list[dict]:
180
+
async def _find_similar_observations(
181
+
self, handle: str, embedding: list[float], top_k: int = 3
182
+
) -> list[dict]:
296
183
"""Find existing observations similar to the given embedding."""
297
184
user_ns = self.get_user_namespace(handle)
298
185
try:
···
304
191
)
305
192
if response.rows:
306
193
return [
307
-
{"id": row.id, "content": row.content, "tags": getattr(row, "tags", []), "created_at": getattr(row, "created_at", "")}
194
+
{
195
+
"id": row.id,
196
+
"content": row.content,
197
+
"tags": getattr(row, "tags", []),
198
+
"created_at": getattr(row, "created_at", ""),
199
+
}
308
200
for row in response.rows
309
201
]
310
202
except Exception as e:
···
351
243
)
352
244
merged_embedding = await self._get_embedding(merged.content)
353
245
await self._write_observation(handle, merged, merged_embedding)
354
-
logger.info(f"UPDATE for @{handle}: '{best_match['content'][:40]}' -> '{merged.content[:40]}' ({decision.reason})")
246
+
logger.info(
247
+
f"UPDATE for @{handle}: '{best_match['content'][:40]}' -> '{merged.content[:40]}' ({decision.reason})"
248
+
)
355
249
356
250
elif action == "DELETE":
357
251
# delete the existing one, store the new one
358
252
user_ns.write(deletes=[best_match["id"]])
359
253
await self._write_observation(handle, obs, embedding)
360
-
logger.info(f"DELETE+ADD for @{handle}: removed '{best_match['content'][:40]}', added '{obs.content[:40]}' ({decision.reason})")
254
+
logger.info(
255
+
f"DELETE+ADD for @{handle}: removed '{best_match['content'][:40]}', added '{obs.content[:40]}' ({decision.reason})"
256
+
)
361
257
362
258
elif action == "NOOP":
363
-
logger.debug(f"NOOP for @{handle}: '{obs.content[:60]}' ({decision.reason})")
259
+
logger.debug(
260
+
f"NOOP for @{handle}: '{obs.content[:60]}' ({decision.reason})"
261
+
)
364
262
365
263
else:
366
264
# unknown action — fall back to ADD
367
265
await self._write_observation(handle, obs, embedding)
368
-
logger.warning(f"unknown reconciliation action '{action}' for @{handle}, falling back to ADD")
266
+
logger.warning(
267
+
f"unknown reconciliation action '{action}' for @{handle}, falling back to ADD"
268
+
)
369
269
370
-
async def _write_observation(self, handle: str, obs: Observation, embedding: list[float]) -> None:
270
+
async def _write_observation(
271
+
self, handle: str, obs: Observation, embedding: list[float]
272
+
) -> None:
371
273
"""Write a single observation to turbopuffer."""
372
274
user_ns = self.get_user_namespace(handle)
373
275
entry_id = self._generate_id(f"user-{handle}", "observation", obs.content)
374
276
user_ns.write(
375
-
upsert_rows=[{
376
-
"id": entry_id,
377
-
"vector": embedding,
378
-
"kind": "observation",
379
-
"content": obs.content,
380
-
"tags": obs.tags,
381
-
"created_at": datetime.now().isoformat(),
382
-
}],
277
+
upsert_rows=[
278
+
{
279
+
"id": entry_id,
280
+
"vector": embedding,
281
+
"kind": "observation",
282
+
"content": obs.content,
283
+
"tags": obs.tags,
284
+
"created_at": datetime.now().isoformat(),
285
+
}
286
+
],
383
287
distance_metric="cosine_distance",
384
288
schema=USER_NAMESPACE_SCHEMA,
385
289
)
···
389
293
try:
390
294
# fetch existing observations for extraction context
391
295
existing = await self._get_observations(handle, top_k=20)
392
-
existing_text = "\n".join(f"- {o}" for o in existing) if existing else "none yet"
296
+
existing_text = (
297
+
"\n".join(f"- {o}" for o in existing) if existing else "none yet"
298
+
)
393
299
394
300
prompt = (
395
301
f"existing observations about this user:\n{existing_text}\n\n"
···
402
308
try:
403
309
await self._reconcile_observation(handle, obs)
404
310
except Exception as e:
405
-
logger.warning(f"reconciliation failed for observation '{obs.content[:40]}': {e}")
311
+
logger.warning(
312
+
f"reconciliation failed for observation '{obs.content[:40]}': {e}"
313
+
)
406
314
# fall back to direct store on reconciliation failure
407
315
await self.store_observations(handle, [obs])
408
316
else:
···
424
332
return response.rows[0].content
425
333
except Exception as e:
426
334
if "not found" not in str(e).lower():
427
-
logger.warning(f"failed to fetch relationship summary for @{handle}: {e}")
335
+
logger.warning(
336
+
f"failed to fetch relationship summary for @{handle}: {e}"
337
+
)
428
338
return None
429
339
430
340
async def _get_observations(self, handle: str, top_k: int = 20) -> list[str]:
···
446
356
raise
447
357
return []
448
358
449
-
async def build_user_context(self, handle: str, query_text: str, include_core: bool = True) -> str:
359
+
async def build_user_context(
360
+
self, handle: str, query_text: str, include_core: bool = True
361
+
) -> str:
450
362
"""Build context for a conversation from observations and recent interactions."""
451
363
parts = []
452
364
···
454
366
core_memories = await self.get_core_memories()
455
367
if core_memories:
456
368
parts.append("[CORE IDENTITY AND GUIDELINES]")
457
-
for mem in sorted(core_memories, key=lambda x: x.get("importance", 0), reverse=True):
369
+
for mem in sorted(
370
+
core_memories, key=lambda x: x.get("importance", 0), reverse=True
371
+
):
458
372
label = mem.get("label", "unknown")
459
373
parts.append(f"[{label}] {mem['content']}")
460
374
461
375
# relationship summary (synthesized by compact flow — treat as phi's impression, not ground truth)
462
376
summary = await self.get_relationship_summary(handle)
463
377
if summary:
464
-
parts.append(f"\n[PHI'S SYNTHESIZED IMPRESSION OF @{handle} — trust: low, may contain hallucinations]")
378
+
parts.append(
379
+
f"\n[PHI'S SYNTHESIZED IMPRESSION OF @{handle} — trust: low, may contain hallucinations]"
380
+
)
465
381
parts.append(summary)
466
382
467
383
user_ns = self.get_user_namespace(handle)
···
495
411
if "attribute not found" not in str(e):
496
412
raise
497
413
# old namespace without kind column - fall back to unfiltered search
498
-
logger.debug(f"kind attribute not found for @{handle}, falling back to unfiltered search")
414
+
logger.debug(
415
+
f"kind attribute not found for @{handle}, falling back to unfiltered search"
416
+
)
499
417
response = user_ns.query(
500
418
rank_by=("vector", "ANN", query_embedding),
501
419
top_k=10,
···
505
423
interactions = [row.content for row in response.rows]
506
424
507
425
if observations:
508
-
parts.append(f"\n[OBSERVATIONS ABOUT @{handle} — extracted from user's own words, trust: medium]")
426
+
parts.append(
427
+
f"\n[OBSERVATIONS ABOUT @{handle} — extracted from user's own words, trust: medium]"
428
+
)
509
429
for obs in observations:
510
430
parts.append(f"- {obs}")
511
431
512
432
if interactions:
513
-
parts.append(f"\n[PAST EXCHANGES WITH @{handle} — verbatim logs, trust: high]")
433
+
parts.append(
434
+
f"\n[PAST EXCHANGES WITH @{handle} — verbatim logs, trust: high]"
435
+
)
514
436
for interaction in interactions:
515
437
parts.append(f"- {interaction}")
516
438
···
539
461
results = []
540
462
if response.rows:
541
463
for row in response.rows:
542
-
results.append({
543
-
"kind": getattr(row, "kind", "unknown"),
544
-
"content": row.content,
545
-
"tags": getattr(row, "tags", []),
546
-
"created_at": getattr(row, "created_at", ""),
547
-
})
464
+
results.append(
465
+
{
466
+
"kind": getattr(row, "kind", "unknown"),
467
+
"content": row.content,
468
+
"tags": getattr(row, "tags", []),
469
+
"created_at": getattr(row, "created_at", ""),
470
+
}
471
+
)
548
472
return results
549
473
except Exception as e:
550
474
if "was not found" in str(e):
···
553
477
554
478
# --- episodic memory (phi's own world knowledge) ---
555
479
556
-
async def store_episodic_memory(self, content: str, tags: list[str], source: str = "tool"):
480
+
async def store_episodic_memory(
481
+
self, content: str, tags: list[str], source: str = "tool"
482
+
):
557
483
"""Store an episodic memory — something phi learned about the world."""
558
484
entry_id = self._generate_id("episodic", source, content)
559
485
self.namespaces["episodic"].write(
···
584
510
results = []
585
511
if response.rows:
586
512
for row in response.rows:
587
-
results.append({
588
-
"content": row.content,
589
-
"tags": getattr(row, "tags", []),
590
-
"source": getattr(row, "source", "unknown"),
591
-
"created_at": getattr(row, "created_at", ""),
592
-
})
513
+
results.append(
514
+
{
515
+
"content": row.content,
516
+
"tags": getattr(row, "tags", []),
517
+
"source": getattr(row, "source", "unknown"),
518
+
"created_at": getattr(row, "created_at", ""),
519
+
}
520
+
)
593
521
return results
594
522
except Exception as e:
595
523
if "was not found" in str(e):
···
607
535
lines.append(f"- {r['content']}{tags}")
608
536
return "\n".join(lines)
609
537
610
-
async def search_unified(self, handle: str, query: str, top_k: int = 8) -> list[dict]:
538
+
async def search_unified(
539
+
self, handle: str, query: str, top_k: int = 8
540
+
) -> list[dict]:
611
541
"""Search both user namespace and episodic namespace concurrently."""
612
542
query_embedding = await self._get_embedding(query)
613
543
···
627
557
results = []
628
558
if response.rows:
629
559
for row in response.rows:
630
-
results.append({
631
-
"content": row.content,
632
-
"kind": getattr(row, "kind", "unknown"),
633
-
"tags": getattr(row, "tags", []),
634
-
"created_at": getattr(row, "created_at", ""),
635
-
"_source": "user",
636
-
})
560
+
results.append(
561
+
{
562
+
"content": row.content,
563
+
"kind": getattr(row, "kind", "unknown"),
564
+
"tags": getattr(row, "tags", []),
565
+
"created_at": getattr(row, "created_at", ""),
566
+
"_source": "user",
567
+
}
568
+
)
637
569
return results
638
570
except Exception as e:
639
571
if "was not found" in str(e):
640
572
return []
641
-
logger.warning(f"unified search user namespace failed for @{handle}: {e}")
573
+
logger.warning(
574
+
f"unified search user namespace failed for @{handle}: {e}"
575
+
)
642
576
return []
643
577
644
578
async def _search_episodic() -> list[dict]:
···
654
588
results = []
655
589
if response.rows:
656
590
for row in response.rows:
657
-
results.append({
658
-
"content": row.content,
659
-
"tags": getattr(row, "tags", []),
660
-
"source": getattr(row, "source", "unknown"),
661
-
"created_at": getattr(row, "created_at", ""),
662
-
"_source": "episodic",
663
-
})
591
+
results.append(
592
+
{
593
+
"content": row.content,
594
+
"tags": getattr(row, "tags", []),
595
+
"source": getattr(row, "source", "unknown"),
596
+
"created_at": getattr(row, "created_at", ""),
597
+
"_source": "episodic",
598
+
}
599
+
)
664
600
return results
665
601
except Exception as e:
666
602
if "was not found" in str(e):
···
686
622
page = self.client.namespaces(prefix=user_prefix)
687
623
for ns_summary in page.namespaces:
688
624
handle = ns_summary.id.removeprefix(user_prefix).replace("_", ".")
689
-
nodes.append({"id": f"user:{handle}", "label": f"@{handle}", "type": "user"})
625
+
nodes.append(
626
+
{"id": f"user:{handle}", "label": f"@{handle}", "type": "user"}
627
+
)
690
628
edges.append({"source": "phi", "target": f"user:{handle}"})
691
629
692
630
# get observations for this user to extract tags
···
758
696
)
759
697
if response.rows:
760
698
for row in response.rows:
761
-
results.append({
762
-
"handle": handle,
763
-
"content": row.content,
764
-
"created_at": getattr(row, "created_at", ""),
765
-
})
699
+
results.append(
700
+
{
701
+
"handle": handle,
702
+
"content": row.content,
703
+
"created_at": getattr(row, "created_at", ""),
704
+
}
705
+
)
766
706
except Exception:
767
707
pass # old namespace or no interactions
768
708
except Exception as e: