Roadmap

A cognitive progression,
built in order.

Storage → episodic memory → semantic memory → collective memory → predictive cognition. Each stage adds a capability on top of the same append-only symbolic substrate. We label maturity honestly, and we don't promise dates.

  1. Shipped · v1

    Episodic memory

    A robust local memory substrate for AI assistants — long-term episodic memory that survives sessions, workspaces, and restarts.

    • Local PostgreSQL + pgvector store, MCP-first
    • Five-axis salience, sleep-like replay and consolidation, controlled forgetting
    • Graph-aware, explainable retrieval; multi-host compatibility

  2. Shipped · v2

    Self-hosted team deployment

    The same cognitive substrate as a secure, multi-tenant deployment you run yourself.

    • Multi-tenancy with row-level security and tenant isolation
    • Hosted MCP, persona-aware retrieval, semantic promotion pipeline
    • Cognition unchanged from v1 — infrastructure separated from the memory model

  3. Shipped · v3

    Self-model layer

    A persistent, governed self-model on top of the episodic and semantic layers.

    • Procedural self-model: behavioral rules and safety constraints (policy_*), agent identity and role boundaries (identity_*)
    • Semantic promotion and sleep-based consolidation hardened in production use

  4. Shipped · v4 · release candidate

    Cognitive governance

    A governance layer over the memory itself: not just who may read the database, but what the system is allowed to do with a given memory. Built entirely as append-only overlays, so the authoritative store is never rewritten.

    • A use-class ladder — a memory is evidence by default, and becomes instruction-grade only through an explicit human confirmation
    • One fail-closed gate on every read surface, with an off / shadow / enforce rollout so it can be observed before it ever denies
    • Deeper sleep phases: typed reasoning edges, interference-based decay, a revision settler that suppresses rather than rejects, and bounded coexisting standpoints for genuine disagreement
    • Operator verbs for tenant erasure, subject-level erasure, and provenance drill-down
    Honest status. This is a release candidate. The read gate ships off by default and our own production deployment runs it off; enforce is still blocked on an ownership-granularity decision. Sensitivity/retention enforcement and provenance drill-down are Pro features, and drill-down additionally requires an explicit operator opt-in.

  5. Not shipped · predictive recall · three sealed NO-GOs

    Predictive recall

    The plan was a latent representation-learning layer (JEPA) enriching retrieval without modifying the symbolic substrate. We pre-registered the bar, ran it, and it did not clear — three times. We record that here rather than quietly move the goalposts.

    • Proof run — NO-GO. The trained predictor scored below both of its own control baselines, and both shortcut probes failed.
    • Objective falsification — hypothesis dead. We pre-registered "the training objective is the problem", tested it, and both arms landed under the bar. The objective was not the problem.
    • Boundary phase — sealed NO-GO. The checkpoint was dropped and a cheap control promoted in its place.

    What survives is narrower and honest: the substrate derives a surprise signal (prediction error, not rankings) during sleep. In this release it is observable but inert — every consumer of it is separately gated. Recall today runs three streams: vector, lexical, and graph.

  6. Planned

    Collective memory

    Let multiple Cogitent instances learn together while each keeps ownership of its local memories — privacy-preserving federation.

    • Memories never leave their owner; only model updates are shared and aggregated
    • Recursive cognitive layers (individual → team → organization), each with its own semantic memory and privacy policy
    • Reusable organizational lessons that retain provenance, confidence, and supporting evidence
    This was previously listed as the v4 milestone. It did not ship in v4 — v4 delivered cognitive governance instead. The aggregator remains a drafted contract and an explicitly fenced skeleton. The governance layer above is the precondition for federation, not a substitute for it.

  7. Vision · v5

    Predictive cognition

    Turn memory from a retrospective store into a predictive engine — anticipating future semantic states through internal simulation during sleep, detecting emerging concepts and knowledge gaps, and navigating memory as a structured space.

    This is a research direction, stated as vision — not a committed release.

How to read this. "Shipped" means the code is built, merged, and running — for v1–v3 that means in production use today; v4 is a release candidate, so it is labelled as one. Where a shipped capability is gated, off by default, or paid, it says so on the entry itself. "Not shipped" means exactly that — including for things we tried and measured as failures, which stay on this page. "Planned" and "vision" are directions, not dated commitments. We move a capability forward only when the evidence supports it, and we move it back when the evidence doesn't.
See the benchmarks → Get launch updates