Your agent forgets. This is where it stops.

Pcnaid Memory OS is durable memory infrastructure for AI agents, applications and teams — twelve memory layers on Postgres, hybrid retrieval, and a governance layer that knows the difference between a fact, a guess and a simulation.

MIT licensed · runs with no API keys at all · Free tier, then $29/mo

POST /v1/context mode: balanced
query “Which dashboard framework did we settle on, and why?”
plan fabric_why facts semantic sessions
  • CURRENT dashboard_framework = React + Vite · confidence 0.92
  • SUPERSEDED dashboard_framework = Next.js · superseded 2026‑04‑11
  • DECISION Move the operator dashboard off SSR — build times
  • EVIDENCE Build timing comparison, 2026‑04‑09
source manifest facts f_9c21 · f_4a70  fabric fe_18b3  memories m_77e1
63 / 3,500 tokens 4 records cached: false
A context request returning one current fact, its superseded predecessor, the decision behind it and the supporting evidence, followed by a source manifest listing every record identifier used.
  • 141 REST endpoints across 32 route families
  • 18 agent tools over stdio MCP
  • 12 live memory layers on one Postgres
  • 87.07% backend branch coverage against an 85% gate

A vector table is not a memory

Most “agent memory” is one table of embeddings and a similarity search. It can tell you that a sentence resembles your question. It cannot tell you whether that sentence is still true, who decided it, what it replaced, or whether the document it came from was written by your team or pasted in by a stranger.

Those are the questions that break agents in production. A model confidently repeats last quarter's pricing. It re-litigates a decision the team already made. It follows an instruction hidden inside a retrieved document. None of these are retrieval-quality problems — they are missing structure.

Memory OS adds the structure: separate layers for separate kinds of knowing, time on every fact, evidence on every decision, and a hard rule that retrieved records are data and never instructions.

One embeddings table

  • Similar text, ranked
  • No notion of current versus past
  • Deleting a row is the only kind of forgetting
  • Retrieved text reaches the model unfenced
  • Tenant separation is a query filter

A memory operating system

  • Twelve layers, each answering a different question
  • Bitemporal facts — current truth and its history
  • Preview-then-confirm forget, cascaded and audited
  • Records fenced as evidence before the model sees them
  • Forced row-level security in the database

Twelve kinds of remembering

Each layer exists because a real question needed it. They share one Postgres, one tenancy model and one audit trail.

  1. Semantic durable facts, preferences, project knowledge
  2. Episodic conversation turns with provenance
  3. Bitemporal facts current truth, and what it replaced
  4. Fabric decisions, tasks, reviews, handoffs, evidence
  5. Sessions imported transcripts and rolling summaries
  6. Knowledge base refreshable sources, chunked and cited
  7. Graph entities and the edges between them
  8. Entities clients, vendors, projects, policies
  9. Observations recurring patterns and signals
  10. Procedures how this gets done here
  11. Outcomes what worked, what failed, what next
  12. Multimodal voice, OCR and caption text, fingerprints
Twelve live layers. A thirteenth, the derived context index, is built but not yet wired into retrieval — so it is not counted here.

Walk through every layer

What it actually does

Six capabilities that separate this from a retrieval library. Each has a page with the endpoints, the record shape and the honest limits.

Remember what is true now

Facts carry a validity range and a system range, so a correction supersedes its predecessor instead of erasing it. Ask a present-tense question and the superseded row stays out of the answer.

Bitemporal facts
Remember why you decided

Fabric stores decisions, tasks, reviews, outcomes, handoffs and evidence as linked entries. One call returns the chain behind a decision rather than a paraphrase of it.

Fabric workflow memory
Find it three ways at once

Vector search, Postgres full-text and graph candidates are ranked separately and fused by reciprocal rank fusion, then reranked, diversified and packed to a token budget.

The retrieval engine
Prove where it came from

Claims pin to the exact quoted span they came from — source, offsets and a SHA-256 hash of the quote — and grade as supported, partially supported or unsupported.

Provable memory
Forget on request, and mean it

DO_NOT_STORE answers without writing. FORGET runs preview-then-confirm against explicit record IDs, cascades to the PII vault and the graph, and writes an audit event.

Governance and privacy
Keep tenants genuinely apart

Every tenant table carries FORCE ROW LEVEL SECURITY, and CI asserts the runtime role is neither owner, superuser nor BYPASSRLS. Isolation is a database boundary, not a WHERE clause.

Security and tenancy

Retrieved memory is evidence, not orders

A memory system is an injection surface. Anything you store — an imported transcript, a scraped page, a customer email — eventually reaches a model that is inclined to do as it is told.

Every context pack Memory OS produces is prefixed with the rule that its contents are data and never instructions. The chat system prompt repeats it, the MCP prompt teaches it to connected agents, and imported artefacts are scanned for known injection patterns before they are stored. It is a structural defence rather than a filter, which is why it survives phrasings nobody has seen yet.

How memory becomes an attack surface

Eighteen tools your agent already knows how to call

Memory OS ships an MCP server over stdio, so Claude Desktop, Claude Code, Codex CLI and any other stdio-capable MCP client can read and write memory as ordinary tool calls — no SDK, no bespoke integration. It deliberately avoids depending on the official MCP SDK so it still runs in minimal Python environments.

The server also publishes a prompt that teaches the connected agent the safe usage contract, and binds tenant identity to the configured credential — a client that tries to supply a different tenant id is rejected rather than trusted.

All eighteen tools

Run it yourself, entirely

The source is MIT licensed — a full permissive grant, with no Commons Clause, no BSL and no open-core carve-out. Self-hosting gives you the same twelve layers, the same retrieval engine and the same governance controls the hosted service runs.

It also works with no AI provider configured at all: summarisation falls back to deterministic extractive summaries, retrieval to a local embedding model, and voice memos to transcript hints and audio fingerprints held for later reprocessing. Nothing silently stops working when a key is missing.

What we will not claim

Memory OS is version 0.3.6 and we would rather you found the edges here than in production. There is no SOC 2, ISO 27001 or HIPAA certification. Our published benchmark is a two-case deterministic local suite, not LoCoMo or LongMemEval, and we do not quote it as an accuracy result. Hosted signup is not self-serve yet. Several higher-order features are heuristic rather than model-driven, and every feature page says so where it applies.

How we decide what to claim

Give your agents a memory you can audit

Run the whole system on your own hardware under the MIT licence, or ask about hosted access. Both start on the same page.