fact_type misclassified — user preferences/rules/corrections stored as "experience" instead of "world"

[jasstionzyf]

Bug Description

Title:

Bug Description

When users state preferences, rules, or corrections during assistant interactions (e.g., "优先用 browser_navigate 而非 web_search"), the retain pipeline classifies them as experience ("assistant") type facts instead of world type.

Root Cause

The LLM extraction prompt in fact_extraction.py uses a rule that is too broad:

> fact_type: use "world" unless the content is clearly an interaction with the assistant.

This causes any fact stated within an interaction to be classified as "assistant" (→ experience), regardless of its nature. User preferences and permanent rules are objective facts about the world — they would exist without this conversation — but are being stored as transient experiences.

Location

hindsight-api-slim/hindsight_api/engine/retain/fact_extraction.py:

- Line 769: "fact_type: use \"world\" unless the content is clearly an interaction with the assistant."
- Lines 1393-1402: raw "assistant" → "experience", "world" → "world" mapping
- Lines 195-196, 298-299, 348-349: Field descriptions that don't distinguish between "doing something" and "saying something"

See also retain/prompts.py line 870:
> world: User's life, other people, events (would exist without this conversation)

This definition is actually correct — the problem is the line 769 rule overrides it.

Consequence

- User preferences stored as experience degrade recall ranking (experience is for one-time events, not persistent rules)
- Consolidation treats them as episodic rather than durable knowledge
- A planned world-fact conflict detection feature would miss them entirely

Proposed Fix

Update the fact_type boundary guidance to distinguish "action" from "statement":

diff
- fact_type: use "world" unless the content is clearly an interaction with the assistant.
+ fact_type 判定：
+ - "world": user preferences, rules, corrections, constraints, and objective facts —
+   even when stated during conversation.
+   Examples: "user wants browser_navigate prioritized", "user forbids unapproved code changes"
+ - "assistant" (→ experience): actions the assistant/agent actually performed,
+   not things discussed.
+   Examples: "I deployed X", "I debugged Y"


Also update the Field descriptions in the Pydantic models to emphasize that factual content remains "world" even when communicated through interaction.

Environment

- Hindsight v0.8.3 (ee81c65)
- LLM: deepseek-v4-flash
- Bank: hermes

Steps to Reproduce

above

Expected Behavior

above

Actual Behavior

above

Version

v0.8.3

LLM Provider

OpenAI