Memory Conflict Detection

The practice of detecting when a new piece of knowledge contradicts an existing one and keeping both visible rather than silently overwriting. This matters for any knowledge system where facts change over time — agent memory, wikis, configuration databases — because silent overwrites destroy the context about what changed and why.

The problem with overwriting

The simplest approach to updating knowledge: find the old entry, replace it with the new one. This works when the new information is unambiguously correct. It breaks when:

• The "new" information might be wrong (an agent misread something, a source is unreliable)
• The old and new information are both partially true in different contexts
• Understanding the contradiction itself is valuable (it reveals a change in the system or a gap in understanding)

In agent memory specifically, silent overwriting means the agent loses history. If a deployment script changed from requiring --force to not requiring it, both facts are useful: the current state and the knowledge that it used to be different.

How conflict detection works

When storing a new memory:

1. Check existing memories for semantic overlap (same topic, similar tags, matching entities)
2. Compare claims — if the new memory asserts something that contradicts an existing memory, flag it
3. Store both memories and create a conflict record linking them
4. Surface conflicts to the agent or human for explicit resolution

The detection can be simple (exact tag + entity match with differing values) or sophisticated (embedding similarity above threshold with contradictory sentiment). hippo-memory uses the simpler approach: conflicts are stored in SQLite and mirrored to .hippo/conflicts/ as files, resolved explicitly with hippo resolve <id> --keep <mem_id>.

Resolution strategies

• Keep newest — trust the most recent observation. Simple but loses history.
• Keep both with context — annotate both with timestamps and let the consumer decide. Preserves the most information.
• Merge — create a new entry that captures the nuance ("deploy required --force before v2.3, no longer needed after"). Most accurate but requires understanding.
• Human arbitration — flag for manual review. Safest for high-stakes knowledge. This is a form of human-in-the-loop applied to knowledge management.

Relevance to this wiki

This wiki handles conflicts through the lint workflow: periodic scans that flag contradictions between pages for human review. The approach is similar in spirit to conflict detection but batch-oriented rather than real-time. If the wiki's ingest rate increases, real-time conflict detection during ingest (checking new claims against existing pages) would catch issues earlier.

proposition-identity-memory-tool is a report from someone who tried to automate step 2, comparing claims, and found that deciding whether two statements say the same thing is proposition identity, an open problem in the philosophy of language. That is an argument for keeping resolution with a human, as the lint workflow does, rather than trusting an automatic contradiction checker.

byzantine-fault detection in multi-agent systems is a related problem at a different scale: agents producing conflicting outputs from the same prompt. The same principle applies — keep contradictions visible and resolve explicitly rather than letting one silently win.