Durable, temporal, governed memory for AI agents.
Longmemory is a local-first, self-hosted memory engine for LLM applications — not just a vector store. It tracks what's true now, what used to be true, and why, across four distinct recall modes.
[ 01 / 04 ]
[ hero ][ 02 / 04 ]
[ features ][ 01 ] [ IMMUTABLE ]
Hydrograph substrate
Every write lands as an immutable node with executable edges, worlds, and entities, not a flat row in a vector index.
[ 03 ] [ context ]
Four recall modes, one engine
Strict, historical, associative, and world-grounded recall run against the same substrate, so each query gets the retrieval contract it actually needs.
Strict recall applies temporal, contradiction, and confidence gates for answers you can act on. Historical and world-grounded recall preserve superseded truth and verify against current external evidence, while associative recall follows semantic, lexical, and graph signals across sessions.
[ 05 ] [ ingestion ]
Session porter & connectors
Import history straight from coding agents and chat tools, then keep ingesting documents and events through the same pipeline.
[ 06 ] [ knowledge ]
Project memory & governed assets
Give coding and workflow agents durable project memory, with governed assets that stay consistent across every session.
[ 07 ] [ providers ]
Eight embedding providers
Swap between OpenAI, Gemini, NVIDIA NIM, AWS Bedrock, Ollama, or a local HTTP endpoint without rewriting your pipeline.
[ 04 ] [ storage ]
Local-first, durable storage
SQLite keeps nodes, edges, worlds, and temporal history on disk, so recall survives restarts without a separate database service.
[ 02 ] [ decay ]
Decay and reinforcement
Confidence decays on a deterministic schedule and strengthens on explicit reinforcement, so memory never needs manual tuning.
[ 03 / 04 ]
[ lifecycle ][ 00 ] [ decay engine ]
Memory that ages the way your product needs it to
Confidence decays deterministically, reinforcement is explicit, and every change leaves a provenance trail back to its trigger.
[ 01 ] [ awareness ]
Deterministic decay
Each memory ages on a predictable curve with a configurable floor, so retrieval stays explainable instead of mysterious.
[ 02 ] [ autonomous ]
Explicit reinforcement
Reinforcement events restore strength on demand, driven by real signals from conversations or tool outcomes, never a hidden background job.
[ 03 ] [ attribution ]
Full provenance
Every reinforcement and decay step links back to its trigger, so agents can explain why a memory is still in play.
[ 04 / 04 ]
[ get started ]Get Started in Minutes
Simple, intuitive API that gets you from zero to production-ready memory in just a few lines of code
[ 01 ] [ conversational ]
Conversational AI
Equip chat products with persistent preferences, tone awareness, and long-tail recollection so they sound brilliantly personal.
GetStarted.js
01 // Ingest a memory through the REST API
02 await fetch ( 'http://localhost:3000/v1/ingest' , {
03 method : 'POST' ,
04 headers : { 'Content-Type' : 'application/json' },
05 body : JSON . stringify ({
06 user_id : 'alice' ,
07 text : 'I prefer TypeScript for backend services'
08 })
09 });
10
11 // Recall with an explicit mode: strict, historical, associative, or world_grounded
12 const response = await fetch ( 'http://localhost:3000/v1/recall' , {
13 method : 'POST' ,
14 headers : { 'Content-Type' : 'application/json' },
15 body : JSON . stringify ({
16 text : 'What language does Alice prefer?' ,
17 mode : 'strict'
18 })
19 });
20
21 const { results } = await response . json ();
22
23 // Results include evidence, confidence, and a traceable provenance chain
[ 02 ] [ assistants ]
Personal assistants
Launch companion agents that accumulate rituals, goals, and collaborative context while staying effortlessly organized.
Longmemory
An open-source project licensed under the Apache License 2.0.
(c) 2026 Longmemory, CaviraOSS and contributors.