A multi-layer agent memory manager modeled after human memory architecture: short-term, mid-term, and long-term stores for persistent, queryable, lifecycle-aware AI memory. The v1 release adds the concrete systems work: hybrid retrieval, live-row invalidation, bi-temporal queries, and a direct retrieve→answer evaluation path.
{card.body}
Humans keep focused working memory, compress recent context, and preserve durable knowledge for semantic recall. Continuum uses the same practical split, then adds database semantics that human memory does not have: supersession, transaction time, and reproducible retrieval.
{body}
session.process_turn(msg) is the only entry-point a caller ever needs. Four steps happen on the foreground path under a strict latency budget. Four more run on the background queue once the reply is already in the caller's hands.
The diagram is a picture; this is the thing. Select one of the prepared statements about location, work, pets, or vehicles. A local in-browser extractor stands in for Continuum's FactExtractor, atomic facts land in the LTM table, and contradictions stamp old rows with invalidated_at as you go.
The diagram below traces a single process_turn() from the user message to the LLM reply, and then continues into the background queue where promotion writes new MTM blocks and LTM facts. Auto-plays — pause, step, or click any phase chip.
The retriever is the only component that touches every tier. Four signal channels are fused into one ranking by the composite scorer — the same scorer that gives the recency channel an honest decay function instead of a hack.
BAAI/bge-reranker-v2-m3) is the most expensive stage and the one that lifts precision the most.
Routing retrieval through three stores instead of one isn't a clever trick — it's an acknowledgement that the value density of memory differs by age. Raw turns are bulky and rarely useful; extracted facts are tiny and answer most queries.
Five strategies in order, each yielding control the moment the bundle fits the caller's token budget. The expensive ones (LLM-as-summarizer, token-level llmlingua) only run when the cheap ones can't.
Promotion is the moment Continuum stops looking like a buffer and starts looking like infrastructure. Three things happen in series, all on the background queue.
Every LTM row carries transaction-time state. invalidated_at TIMESTAMPTZ NULL is the hot-path live guard: NULL means "this fact is current"; a timestamp means the row is historical.
At read-time the retriever adds one clause —{' '}
WHERE invalidated_at IS NULL
{' '}— and gets correct-by-construction "what is current" queries. The same table still serves audit and history reads when that clause is dropped.
Two timestamp columns on every LTM row.
valid_from is the point in real-world time when the fact became true.
recorded_at is when Continuum learned about it.
They usually match — facts are recorded right after they become true — but the system stays correct when they diverge: point-in-time queries, retroactive corrections, late-arriving evidence.
valid_from ≤ 2024-03.
valid_from to the historical date but recorded_at to today.
The publishable v1 story is LongMemEval-S 60.8% judged on 500 questions, plus the memory-operation benchmarks that prove the schema.{' '}
make bench-all reproduces the local memory table; make repro-everything reproduces the v1 headline with an OpenRouter key.
The May report predicted the 32-34% ceiling would need agentic iterative reasoning. v1 built that loop, tested it, and cut it. The shipped result is deliberately simpler: retrieve the right context once, then let a strong answerer answer.
The negative result remains useful. continuum/reasoning/iterative_reasoner.py is still in-tree, tested, and documented, but the public v1 pipeline does not depend on it.
--reasoner direct hands raw retrieved turns to the answerer in one call. This is the v1 winner.
One user turn through a healthy Continuum deployment. Every box below is a swappable component behind a Protocol — the defaults are what the benchmarks use.
{r.a}
No Postgres needed for the demo or the four local memory-operation benchmarks. An OpenRouter key is only required if you want to reproduce the v1 LongMemEval-S headline.
{cmd}
{citation}