Illustration of a writer at a desk with memories flowing from thought into books and time

Long-term memory for agents

Turn conversations into durable context

Recollect extracts atomic facts from chat, scopes them per user or session, and surfaces the right memories when your agent needs them.

View source Interactive Showcase → Quickstart

Memory that outlives the context window

Agents forget. Preferences, constraints, and confirmed outcomes disappear when the session ends. Recollect gives you a small, embeddable library to add facts from dialogue and search them later with semantic, keyword, and entity signals—without locking you into a hosted platform.

How it works

  1. 1

    Ingest

    Pass raw text or full message threads. Optional LLM extraction yields ADD-only atomic facts.

  2. 2

    Scope

    Tag memories with user_id, agent_id, or run_id filters.

  3. 3

    Embed & store

    Vectors and metadata land in local SQLite under ~/.recollect by default.

  4. 4

    Retrieve

    Hybrid ranking blends embeddings, BM25-style keywords, and entity overlap for the top hits.

Built for production-minded prototypes

ADD-only extraction

Facts accumulate over time. Nothing is silently overwritten during ingestion.

Hybrid retrieval

Semantic similarity plus keyword and entity boosts in one fused score.

Local-first

Ship with SQLite and a deterministic local embedder for tests—add OpenAI when ready.

Simple API

add, search, get, get_all, delete—five methods to start.

See it in action

Support bots, persistent assistants, and multi-agent scoping—with a fully interactive browser playground on the showcase page (add, search, and chat loop all run live).

Try the interactive demo

Quickstart

git clone https://github.com/cobusgreyling/recollect.git
cd recollect
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
from recollect import Memory, RecollectConfig
from recollect.config import EmbedderConfig

memory = Memory(
    RecollectConfig(
        extraction_enabled=False,
        embedder=EmbedderConfig(provider="local"),
    )
)

memory.add("Prefers PostgreSQL over MySQL", user_id="alice", infer=False)
print(memory.search("database", filters={"user_id": "alice"}))