Cognitive memory
for AI agents.

A persistent-memory substrate that lets your coding assistant, support bot, or research agent remember context and recall what matters — across sessions, workspaces, and restarts.

Get launch updates Explore the architecture → 
$ pip install cogitent
$ cogitent-mcp serve --transport stdio
# your agent now has long-term memory.

LLMs are powerful but amnesiac. Their context window forgets everything the moment a session ends. Cogitent is the missing substrate: episodic storage, salience-weighted recall, and consolidation cycles inspired by how cognition actually works. It works across MCP-compatible AI hosts and major providers, so your agent keeps its memory whichever model you run.

What's in the box

Persistent recall

Episodes survive across sessions, workspaces, and host restarts. Cache-hit recalls return in <1 ms; semantic recalls in <50 ms p50. Retrieval quality measures 65.0 % LongMemEval-S, fully local (retrieval-quality measurement).

Salience-weighted

Five-axis salience (importance, urgency, novelty, confidence, affect) ranks what surfaces. Background sleep cycles consolidate, decay, and settle revisions — a unique cognition architecture among surveyed memory platforms. Predictive recall was measured against a pre-registered bar and did not clear it; see the roadmap.

Open core, EU-built

AGPL-3.0-or-later Community Edition — free and self-hostable forever. Memory is stored and processed locally; recalled context is shared only with the AI host you choose. French-hosted federated learning is on the roadmap. Plus for AI-intensive individuals & small teams, Pro for enterprises; full self-host always supported.

Governed memory

A memory is evidence by default — it only becomes something your agent treats as an instruction after an explicit human confirmation. Per-memory use-class, sensitivity, and sharing rules are enforced by a single fail-closed gate on every read. The gate ships off by default with a shadow mode, so you can measure its blast radius before it ever denies anything. Sensitivity & retention enforcement is a Pro feature.

Editions

Self-host the core forever. Upgrade when you want team-scale features.

Community Edition

Free · AGPL-3.0-or-later

  • Local-first persistent memory (Postgres + pgvector)
  • MCP server (stdio + HTTP)
  • TUI + CLI
  • All recall modes (cache, semantic, hybrid)
  • Sleep cycles & consolidation
  • Anonymous opt-in federated learning (roadmap)
Get notified at launch
90-day free trial

Cogitent Plus

Pricing at launch · per user / month

  • Everything in CE
  • Production-grade defaults & hardening
  • Governance observability & audit-log viewer available today
  • Coexisting-standpoint conflicts view available today
  • Toggleable telemetry & FL participation at launch
  • Priority updates & support
  • Hosted entitlement & revocation
Get pricing news
Enterprise

Cogitent Pro

Talk to us · annual contract

  • Everything in Plus
  • Sensitivity & retention enforcement available today
  • Tool-output provenance drill-down operator opt-in
  • Team governance: access & sharing controls at launch
  • Enterprise SSO & compliance bundles at launch
  • Fleet deployment at launch
  • Dedicated support SLA
  • Air-gapped federated-learning hosting roadmap
Contact for enterprise

CE is the individual layer, free forever. Organisations buy governance and accountability — never access to their own memory.
Reading the labels: available today means the entitlement is enforced in the current release candidate. at launch and roadmap are not shipped yet and are not billed as if they were.