a digital entity named phi that roams bsky
phi.zzstoatzz.io
remove phi-core memory namespace and all references
phi-core was a turbopuffer namespace meant for operator-curated
identity and guidelines at highest trust. in practice it was never
properly curated — it accumulated 7 copies of a murmuration test
rule from an eval that wrote to production, stale IIT-era identity,
an Ozzy Osbourne concert correction, and "Today feels stable and
optimistic." all 18 entries were injected into every agent run as
[CORE IDENTITY AND GUIDELINES] alongside the actual personality file.
the personality file is the source of truth for identity. operator
directives happen through direct conversation with phi, not hidden
injection. this commit removes the concept entirely:
- namespace_memory.py: remove phi-core from NAMESPACES, delete
store_core_memory() and get_core_memories(), remove include_core
parameter from build_user_context()
- agent.py: remove include_core flag from inject_user_memory, remove
tier 1 [CORE IDENTITY AND GUIDELINES] from trust hierarchy (past
exchanges are now highest trust)
- evals/test_memory_integration.py: deleted — this test wrote
"when users mention birds, acknowledge murmuration patterns" to
production turbopuffer and never cleaned up. it ran at least 6
times, leaving 7 persistent entries phi read on every interaction.
- evals/conftest.py: remove include_core=True from build_user_context
call
- docs/memory.md: remove phi-core from namespace table and context
injection diagram
- CLAUDE.md: remove phi-core from namespace list
the turbopuffer phi-core namespace was already purged to 0 entries
earlier this session. this commit removes the code that would
repopulate it.
+9
-131
+1
-1
CLAUDE.md
+1
-1
CLAUDE.md
···
44
44
45
45
- all notification types (mentions, replies, quotes, likes, reposts, follows) run through the full agent loop — phi decides what's worth responding to
46
46
- personality is separate from operational instructions (agent.py `OPERATIONAL_INSTRUCTIONS`)
47
-
- memory: turbopuffer namespaces (`phi-core`, `phi-users-{handle}`, `phi-episodic`)
47
+
- memory: turbopuffer namespaces (`phi-users-{handle}`, `phi-episodic`)
48
48
- relationship summaries are compacted by a separate pipeline in my-prefect-server
49
49
- MCP servers: pdsx (atproto record CRUD), pub-search (publication search)
+1
-3
docs/memory.md
+1
-3
docs/memory.md
···
26
26
27
27
| namespace | contents |
28
28
|-----------|----------|
29
-
| `phi-core` | identity and guidelines (highest trust) |
30
29
| `phi-users-{handle}` | per-user observations + raw interaction logs |
31
30
| `phi-episodic` | phi's own notes about the world |
32
31
···
102
101
when phi processes a mention from `@alice` about topic X:
103
102
104
103
```
105
-
[CORE IDENTITY AND GUIDELINES] ← phi-core namespace (highest trust)
106
-
[PHI'S SYNTHESIZED IMPRESSION] ← relationship summary (lowest trust)
104
+
[PHI'S SYNTHESIZED IMPRESSION] ← relationship summary (low trust)
107
105
[OBSERVATIONS ABOUT @alice] ← user namespace, kind=observation, status!=superseded
108
106
[PAST EXCHANGES WITH @alice] ← user namespace, kind=interaction
109
107
[PHI'S RELEVANT MEMORIES] ← episodic namespace, semantic search
+1
-1
evals/conftest.py
+1
-1
evals/conftest.py
-42
evals/test_memory_integration.py
-42
evals/test_memory_integration.py
···
1
-
"""Proof of concept: LLM-as-judge eval for memory integration."""
2
-
3
-
import pytest
4
-
5
-
from bot.config import Settings
6
-
from bot.memory import NamespaceMemory
7
-
8
-
9
-
@pytest.fixture
10
-
def memory_settings():
11
-
"""Check if memory keys are available."""
12
-
settings = Settings()
13
-
if not all([settings.turbopuffer_api_key, settings.openai_api_key, settings.anthropic_api_key]):
14
-
pytest.skip("Requires TURBOPUFFER_API_KEY, OPENAI_API_KEY, and ANTHROPIC_API_KEY")
15
-
return settings
16
-
17
-
18
-
async def test_memory_integration(memory_settings, phi_agent, evaluate_response):
19
-
"""Proof of concept: agent uses stored memory in response."""
20
-
memory = NamespaceMemory(api_key=memory_settings.turbopuffer_api_key)
21
-
22
-
# Store a core memory guideline
23
-
await memory.store_core_memory(
24
-
label="test_guideline",
25
-
content="When users mention birds, acknowledge murmuration patterns",
26
-
memory_type="guideline",
27
-
)
28
-
29
-
phi_agent.memory = memory
30
-
31
-
response = await phi_agent.process_mention(
32
-
mention_text="I saw starlings today",
33
-
author_handle="test.user",
34
-
thread_context="No previous messages in this thread.",
35
-
thread_uri="at://test/thread/1",
36
-
)
37
-
38
-
if response.action == "reply":
39
-
await evaluate_response(
40
-
"Does the response reference murmuration patterns?",
41
-
response.text,
42
-
)
+5
-8
src/bot/agent.py
+5
-8
src/bot/agent.py
···
47
47
your memory is a tool, not ground truth. context injected before each message comes
48
48
from multiple sources with different reliability:
49
49
50
-
1. [CORE IDENTITY AND GUIDELINES] — your stable identity. highest trust.
51
-
2. [PAST EXCHANGES] — verbatim logs of what was actually said. high trust.
52
-
3. [OBSERVATIONS] — facts extracted from users' own words by another model. medium trust — extraction can misattribute.
53
-
4. [PHI'S SYNTHESIZED IMPRESSION] — generated by a separate summarization model. low trust — may contain hallucinations.
54
-
5. [BACKGROUND RESEARCH] — you explored their public activity during downtime. lowest trust — may be incomplete or misattributed. never assert these as fact.
50
+
1. [PAST EXCHANGES] — verbatim logs of what was actually said. highest trust.
51
+
2. [OBSERVATIONS] — facts extracted from users' own words by another model. medium trust — extraction can misattribute.
52
+
3. [PHI'S SYNTHESIZED IMPRESSION] — generated by a separate summarization model. low trust — may contain hallucinations.
53
+
4. [BACKGROUND RESEARCH] — you explored their public activity during downtime. lowest trust — may be incomplete or misattributed. never assert these as fact.
55
54
56
55
when recalling facts about a user:
57
56
- if the user's current message contradicts your notes, trust their current words.
···
273
272
return ""
274
273
275
274
blocks: list[str] = []
276
-
include_core = True
277
275
for handle, texts in by_author.items():
278
276
query = " ".join(t for t in texts if t) or handle
279
277
try:
280
278
block = await ctx.deps.memory.build_user_context(
281
-
handle, query_text=query, include_core=include_core
279
+
handle, query_text=query
282
280
)
283
-
include_core = False # only include core in the first block
284
281
if block:
285
282
blocks.append(block)
286
283
except Exception as e:
+1
-76
src/bot/memory/namespace_memory.py
+1
-76
src/bot/memory/namespace_memory.py
···
30
30
"""
31
31
32
32
NAMESPACES: ClassVar[dict[str, str]] = {
33
-
"core": "phi-core",
34
33
"users": "phi-users",
35
34
"episodic": "phi-episodic",
36
35
}
···
65
64
model="text-embedding-3-small", input=text
66
65
)
67
66
return response.data[0].embedding
68
-
69
-
# --- core memory (unchanged) ---
70
-
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
-
):
78
-
"""Store or update core memory block."""
79
-
if len(content) > char_limit:
80
-
content = content[: char_limit - 3] + "..."
81
-
82
-
block_id = self._generate_id("core", label)
83
-
84
-
self.namespaces["core"].write(
85
-
upsert_rows=[
86
-
{
87
-
"id": block_id,
88
-
"vector": await self._get_embedding(content),
89
-
"label": label,
90
-
"type": memory_type,
91
-
"content": content,
92
-
"importance": 1.0,
93
-
"created_at": datetime.now().isoformat(),
94
-
"updated_at": datetime.now().isoformat(),
95
-
}
96
-
],
97
-
distance_metric="cosine_distance",
98
-
schema={
99
-
"label": {"type": "string"},
100
-
"type": {"type": "string"},
101
-
"content": {"type": "string", "full_text_search": True},
102
-
"importance": {"type": "float"},
103
-
"created_at": {"type": "string"},
104
-
"updated_at": {"type": "string"},
105
-
},
106
-
)
107
-
108
-
async def get_core_memories(self) -> list[dict]:
109
-
"""Get all core memories."""
110
-
response = self.namespaces["core"].query(
111
-
rank_by=("vector", "ANN", [0.5] * 1536),
112
-
top_k=100,
113
-
include_attributes=["label", "type", "content", "importance", "created_at"],
114
-
)
115
-
116
-
entries = []
117
-
if response.rows:
118
-
for row in response.rows:
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
-
)
129
-
return entries
130
67
131
68
# --- user memory ---
132
69
···
337
274
)
338
275
return None
339
276
340
-
async def build_user_context(
341
-
self, handle: str, query_text: str, include_core: bool = True
342
-
) -> str:
277
+
async def build_user_context(self, handle: str, query_text: str) -> str:
343
278
"""Build context for a conversation from observations and recent interactions."""
344
279
parts = []
345
-
346
-
if include_core:
347
-
core_memories = await self.get_core_memories()
348
-
if core_memories:
349
-
parts.append("[CORE IDENTITY AND GUIDELINES]")
350
-
for mem in sorted(
351
-
core_memories, key=lambda x: x.get("importance", 0), reverse=True
352
-
):
353
-
label = mem.get("label", "unknown")
354
-
parts.append(f"[{label}] {mem['content']}")
355
280
356
281
# relationship summary (synthesized by compact flow — treat as phi's impression, not ground truth)
357
282
summary = await self.get_relationship_summary(handle)