Choosing an agent memory layer
RE-call compared with other memory layers
These projects solve different parts of agent memory, and several of them are further along than RE-call. The useful question is not which is best but what each one does when a remembered fact stops being true, and what it needs to run.
Descriptions of other projects come from their public repositories and documentation, checked on 2026-11-05. They change quickly; follow the links before relying on a detail. The sourced, per-capability record is the prior art comparison.
Which one fits
| Choose | When you want | What it needs |
|---|---|---|
| RE-call | Memory that tells the agent a result is superseded, shows where it came from, and refuses (ABSTAIN) when nothing in memory supports an answer. | A local SQLite file (Lite) or PostgreSQL with pgvector. No LLM call in the memory layer by default. |
| Mem0 | An LLM that extracts memories from conversations for you, as a library, self-hosted or as a managed cloud service. | An LLM for extraction, plus a vector store. |
| Graphiti (behind Zep) | A temporal knowledge graph whose ingestion infers entities and relations and invalidates contradicted facts. | A graph database (Neo4j, FalkorDB, Kuzu or Amazon Neptune) and LLM calls at ingestion. |
| Letta | A complete stateful agent runtime in which the agent edits its own memory blocks and searches archival memory. | The Letta server, self-hosted or on Letta Cloud. It is an agent layer, not a retrieval library. |
| LangMem | Memory management tools inside a LangGraph application. | Usually LangGraph and its store. |
| Plain pgvector or a vector database | Nearest-neighbour text, where your own code decides whether a result is current or trustworthy. | The database alone. |
What happens to a fact that stopped being true
- RE-call marks it
supersededwhen a newer memory declares that it replaces it, and returns the current one ahead of it. The author declares the replacement; RE-call does not infer it. - Graphiti infers the contradiction with an LLM at ingestion and invalidates the older edge rather than deleting it.
- Mem0, since its 2026 algorithm, adds memories without updating or deleting them on the add path, and handles time at retrieval. Explicit update and delete remain in its API.
- Letta leaves it to the agent, which rewrites its own memory blocks.
- A plain vector store returns the stale text whenever it is the nearest match.
What RE-call gives up
- Coverage of supersession. Declared replacements are precise but rare: on the reference corpus, 2 of 792 memos declared one while 60 closed a decision only in prose. An LLM that infers replacements covers more and is sometimes confidently wrong; RE-call chooses precision.
- No entity reasoning. There is no knowledge graph of people and relations to reason over, which is where Graphiti is stronger.
- Abstention costs answers. Out of the box, default retrieval abstained on none of LoCoMo's 446 adversarial questions. The shipped levers raise that to between 0.37 and 0.77 only by refusing a quarter to half of legitimate questions.
- Lite installs from source for now. The SQLite edition is newer than the published 0.14.0 wheel; see the Lite guide.
The one paired measurement
On LoCoMo's 1,540 answerable questions, with the same GPT-4o reader and judge and matched retrieval budgets, RE-call scored 48.4% judged answer accuracy against Mem0 OSS at 44.4%. RE-call ran this test itself, and the direction depends on the reader: a later Sonnet 4.5 subset reverses it. The method, artifact and limits are on the benchmark scorecard.
RE-call has not been compared head to head with Graphiti, Zep, Letta or LangMem, so this page makes no score claim about them.
Next
- Install RE-call, Lite or PostgreSQL.
- How trusted search and the trust gate work.
- Evidence and limits of persistent memory.