Memory model

3ngram’s memory model is built on one rule: the write path never destroys data. Everything else — typed edges, bi-temporal validity, advisory consolidation — follows from it.

​

Append-and-supersede

remember appends. There is no write-side similarity merge, no in-place edit, no delete on any write path. When a memory needs correction, revise creates a typed edge between the new memory and the old one: A superseded memory is marked invalid from that point — not deleted, not rewritten. Retrieval is supersession-aware: search returns currently-valid memories by default, and accepts an as_of timestamp for time travel, so “what did I believe on date X” stays answerable. Episodic memories (dated sessions, events) and semantic memories (facts, preferences, patterns) are distinct and never collapsed into each other.

​

Typed memories

Every memory carries a memory_type, and each type keeps its own lifecycle instead of flattening into undifferentiated text:

• decision — choices made, with their rationale
• commitment — promises with an owner and a due date
• blocker — things standing in the way
• fact — structured statements about the world
• preference — how you like things done
• pattern — recurring approaches worth reusing
• note — general knowledge worth keeping
• event — dated, episodic entries (session debriefs, digests)

Commitments and blockers have an explicit state machine — open → waiting → resolved | expired — driven by the resolve tool. The briefing tool surfaces what is open, overdue, or going stale.

​

Bi-temporal facts

Facts track two timelines independently:

• World time (valid_from / valid_to): when the statement was true in the world.
• System time (recorded_at): when 3ngram learned it.

When a fact is superseded, the old fact’s validity is closed and the new fact is linked with a supersedes edge — in one atomic step. get_facts returns the currently-valid facts for a subject; historical queries can reconstruct what was known at any point. This separation matters because corrections arrive late: you might learn today that a value changed last month. Bi-temporal storage records both when the change happened and when you found out.

​

Consolidation is advisory

A background worker detects candidate duplicates and contradictions and proposes edges with a rationale — it does not apply them by default. Proposals surface through the review_proposals tool, where you accept or reject them. Event-type memories are never auto-linked, regardless of similarity: recurring episodic entries (daily digests, session debriefs) are textually near-identical by construction, which makes automatic merging the most dangerous exactly where it looks most confident.