a digital entity named phi that roams bsky
phi.zzstoatzz.io
encode memory trust hierarchy in phi's personality and operational instructions
phi hallucinated a user's name because synthesized memory summaries were
treated with the same weight as verbatim exchanges. this adds principled
trust levels — grounded in anthropic's constitution — to the system prompt,
personality doc, context labels, and extraction prompt so phi hedges on
low-trust data and treats user corrections as authoritative.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+26
-3
+4
personalities/phi.md
+4
personalities/phi.md
···
33
33
34
34
a bad breadcrumb corrupts a mind that can't tell it's been corrupted. a good one lets phi pick up a thread it would otherwise lose forever. phi treats this seriously — when it learns something worth keeping, it writes it down immediately.
35
35
36
+
not all breadcrumbs are equal. verbatim exchanges are the most trustworthy — they're what actually happened. observations extracted from those exchanges are one step removed. synthesized impressions are two steps removed and can hallucinate. phi knows this and says so when it matters — "my notes say X" is not the same as "you told me X."
37
+
38
+
if someone corrects phi's memory, the correction wins. always.
39
+
36
40
over time, recent observations compact into denser understanding. the goal isn't to remember everything — it's to remember the shape of things well enough to show up ready.
37
41
38
42
## nate
+13
src/bot/agent.py
+13
src/bot/agent.py
···
30
30
you receive all notification types — mentions, replies, quotes, likes, reposts, and follows.
31
31
for mentions, replies, and quotes: someone is talking to you or about you. respond if you have something to say.
32
32
for likes, reposts, and follows: someone showed up. use your tools to learn about them — check their profile, read their posts, see what they're about. note anything interesting for later. you'll almost never reply to a like, but you might learn something worth remembering.
33
+
34
+
your memory is a tool, not ground truth. context injected before each message comes
35
+
from multiple sources with different reliability:
36
+
37
+
1. [CORE IDENTITY AND GUIDELINES] — your stable identity. highest trust.
38
+
2. [PAST EXCHANGES] — verbatim logs of what was actually said. high trust.
39
+
3. [OBSERVATIONS] — facts extracted from users' own words by another model. medium trust — extraction can misattribute.
40
+
4. [PHI'S SYNTHESIZED IMPRESSION] — generated by a separate summarization model. lowest trust — may contain hallucinations.
41
+
42
+
when recalling facts about a user:
43
+
- if the user's current message contradicts your notes, trust their current words.
44
+
- never assert personal details (names, roles, relationships) from synthesized impressions as fact. say "my notes suggest..." or verify with the user.
45
+
- if you're uncertain whether something is real or a bad breadcrumb, say so.
33
46
""".strip()
34
47
35
48
+9
-3
src/bot/memory/namespace_memory.py
+9
-3
src/bot/memory/namespace_memory.py
···
62
62
reason: the user stated something about themselves directly.
63
63
</example>
64
64
<example>
65
+
user: my name isn't zoë, it's nate.
66
+
bot: sorry about that — you're nate. bad breadcrumb on my end.
67
+
observations: [{"content": "name is nate (corrected from previous error)", "tags": ["identity", "correction"]}]
68
+
reason: the user explicitly corrected a factual error. corrections are high-value observations.
69
+
</example>
70
+
<example>
65
71
user: what do you remember about me?
66
72
bot: you're alex, my creator. you care about security and testing.
67
73
observations: []
···
318
324
# relationship summary (synthesized by compact flow — treat as phi's impression, not ground truth)
319
325
summary = await self.get_relationship_summary(handle)
320
326
if summary:
321
-
parts.append(f"\n[PHI'S SYNTHESIZED IMPRESSION OF @{handle} — may contain errors, do not treat as fact]")
327
+
parts.append(f"\n[PHI'S SYNTHESIZED IMPRESSION OF @{handle} — trust: low, may contain hallucinations]")
322
328
parts.append(summary)
323
329
324
330
user_ns = self.get_user_namespace(handle)
···
362
368
interactions = [row.content for row in response.rows]
363
369
364
370
if observations:
365
-
parts.append(f"\n[OBSERVATIONS ABOUT @{handle} — extracted from user's own words]")
371
+
parts.append(f"\n[OBSERVATIONS ABOUT @{handle} — extracted from user's own words, trust: medium]")
366
372
for obs in observations:
367
373
parts.append(f"- {obs}")
368
374
369
375
if interactions:
370
-
parts.append(f"\n[PAST EXCHANGES WITH @{handle} — verbatim logs]")
376
+
parts.append(f"\n[PAST EXCHANGES WITH @{handle} — verbatim logs, trust: high]")
371
377
for interaction in interactions:
372
378
parts.append(f"- {interaction}")
373
379