Memory for AI agents that have to act — not just chat about a corpus.
Pour documents into it. Get back what sources said, what the system currently holds true, and a full audit trail to the exact span, page, or second of audio. Built to stay useful at a million documents.
Docs: docs.remember.dev · Product: remember.dev
Most “memory” stacks answer: where did I read something like this?
Agents that take real actions need a harder question:
What do we actually know — and what changed our mind?
| Typical RAG / note memory | RememberStack |
|---|---|
| Source text treated as truth | Claims (testimony) stay separate from facts (current belief) |
| Edits overwrite history | Supersession closes a window — history stays queryable |
| Contradictions hidden or averaged | Contradictions return together |
| Re-ingest inflates “confidence” | Support counts independent document lineages |
| Vector index is the authority | Indexes nominate; Postgres confirms |
| LLM on every query | No chat-completion on the query path — the agent plans |
| Vague empty results | Typed negatives: unknown entity / known empty / boundary |
If your agent spends money, changes state, or briefs a human, those distinctions are load-bearing.
- Ingest heterogeneous inputs into an evidence spine (files → chunks → claims → facts).
- Separate what a source said from what is true now.
- Project search, graph, and a browsable filesystem — rebuildable anytime.
- Serve agents first: mounts, MCP, CLI, API — with honest, grain-typed answers.
E what we ingested (ground truth)
K what we concluded (compiled + authored knowledge)
P how we reach it (search · graph · corpus FS) ← always rebuildable from E
| Grain | Answers | Rule for agents |
|---|---|---|
| Evidence (claims) | Who said what, when | Never “is it true now?” |
| Fact (relations & observations) | What we currently hold true | Default for present-tense belief |
| Compiled (knowledge pages) | Orientation with citations | Verify before load-bearing action |
Default reading motion:
Orient on knowledge pages and the corpus tree → verify on facts → audit claims and raw sources when stakes demand it.
Every fact carries world time (when it held in the world) and system time (when this deployment learned it).
Ask both honestly:
- “Who worked at Acme in 2022?”
- “What did we believe last March?”
- Immutable claims, grounded to source spans
- Entity resolution into a canonical registry
- Adjudicated relations and observations with supersession + contradictions
- Document versions and watched sources — reprocess cost proportional to the edit
- Support that cannot be gamed by re-extracting the same file
Deep dive: Ingestion
Projections nominate. The spine confirms. The envelope accounts.
Exactly four top-level assured operations (API / CLI / MCP):
| Operation | Use for |
|---|---|
resolve_entity |
Name → ranked entity candidates |
testimony_context |
High-recall evidence for a question |
fact_context |
Current or historical fact context with live testimony |
answer_context |
Both complete authority views in ContextBundle/v1 |
Plus open SQL, typed live-graph helpers, saved examples, and schema discovery.
Every assured answer self-accounts: grain, freshness, contradictions, truncation, typed “no”s.
Deep dive: Retrieval
| Surface | Job |
|---|---|
| Filesystem mounts | ls / read / grep the corpus and knowledge like a codebase |
| MCP · CLI · API | Semantic search, graph, time-travel, open query — one operation set |
| Consumption skill | Deployment-rendered SKILL.md that keeps grains straight |
Primary consumers are coding harnesses (Claude Code, Codex, OpenCode, and peers). Humans get the same audit trail.
git clone https://github.com/writeitai/remember-stack.git
cd remember-stack
cp .env.example .env # set your OpenRouter (or provider) key
docker compose up --build --detach --wait
curl --fail http://localhost:8000/healthz
curl --fail http://localhost:8000/operationsIngest Markdown, wait for readiness, then call the assured ops — full walkthrough:
→ Getting started
→ Self-host deployment
Client package:
pip install rememberstack
# server / connectors / knowledge extras named in the packageApache-2.0. If it affects correctness, it is here — extraction, resolution, supersession, provenance, budgets, DLQ, hard-forget. Never paywalled.
The managed cloud runs this same engine. Cloud adds operations and product chrome, not a secret core.
| Docs | docs.remember.dev |
| Managed product | remember.dev |
| Release | v0.13.0 |
See CONTRIBUTING.md and CLA.md. Pull requests need the contributor-agreement checkbox in the PR template.
Stop retrieving passages. Start knowing what is true.
Read the docs · Run it