Enterprise · Non-Human Intelligence (NHI) memory substrate

Self-improving · Governed · Bias-resistant

Memory your AI fleet can
trust, prove, and control.

ai-memory turns AI agents from amnesiac chatbots into a coordinated workforce with a shared, governed, auditable institutional memory — what you need before you can trust AI to operate autonomously at scale. One Rust binary. Your infrastructure. Any LLM. Apache-2.0.

The problem

AI assistants have amnesia.

Without memory

Every conversation starts from zero. The AI forgot what you told it yesterday, what it figured out last week, and every decision you made together. Across a team of agents it's worse: they each forget, redo each other's work, and leave no shared record of what any of them knew or did.

With ai-memory

A long-term memory layer underneath your agents — durable, shared, and trustworthy. The difference between an employee who keeps an organized notebook (and can hand it to a teammate) and one who wakes up every morning with no recollection of the job.

For the CISO / CTO room

Eight reasons enterprises care.

Built like infrastructure, not a demo — because autonomous AI at scale is where the value and the risk both live.

01 · PROVENANCE

Cryptographic audit trail

Append-only, hash-chained, Ed25519-signed ledger. Prove who recorded or changed what, and when — tamper-evident. Turns "the AI did something" into a defensible record.

02 · CONTROL

Governance that fails closed

Per-namespace policy (write / promote / delete / approve), signed rules, secure-by-default. A policy-check error refuses the write — no silent bypass.

03 · IDENTITY

Attestation for non-human actors

Real cryptographic agent identities, mTLS, cert-chain delegation, replay protection. Enrol, scope, and revoke AI agents like service accounts.

04 · SOVEREIGNTY

Deploy anywhere, no lock-in

SQLite (single binary, air-gappable, edge) or PostgreSQL + Apache AGE (scale, graph). Works with any LLM / embedding provider, switchable by config. Offline is first-class.

05 · SECURITY

Encryption + secret hygiene

At-rest encryption (SQLCipher), per-memory encrypted envelopes, TLS in transit, SSRF guards. Secrets structurally barred from logs, audit, and capabilities.

06 · SCALE

Multi-agent, not one chatbot

Federation with quorum (W-of-N) agreement before knowledge becomes authoritative, per-agent quotas, and graceful load-shedding admission control.

07 · SAFETY

AI-safety guardrails

Governance gates, attestation, and anti-monoculture controls that make "is our shared AI knowledge being corrupted or dominated?" visible and measurable.

08 · LONGEVITY

Built to last & be audited

Three interfaces (API, CLI, the MCP standard), a versioned wire/schema contract, lossless archive/restore, enforced engineering discipline. A 5-year foundation.

🔏

Prove it

Cryptographic audit trail.

🛡️

Govern it

Fail-closed policy + attested identity.

🏛️

Own it

Your infra, no AI-vendor lock-in.

The wazoo tier

Self-improving, bias-resistant AI cognition.

Most "AI memory" is a notepad. This is a self-improving cognitive substrate for non-human intelligence — it learns from its own learning, governs itself, proves what it knows, and actively fights the echo chamber.

♻️ Recursive learning & self-improvement

  • Reflections are memories about memories — clusters observations into higher-level insights, and reflects on reflections: a tower of abstraction.
  • Depth-capped by design — a governance-enforced recursion ceiling with a signed audit row per refusal. Smarter without runaway self-reference.
  • Autonomous curator daemon — background reflect / consolidate near-duplicates / decay stale confidence / re-classify, all audited and operator-reversible. Improves while no human watches.
  • Confidence calibration + decay — beliefs lose authority over time unless reinforced. Knowledge ages like a mind, not a static dump.

🧩 Memory skills — learning becomes capability

  • Register, retrieve, compose, and export reusable skills across the whole fleet.
  • Promote-from-reflection: a hard-won insight becomes a callable, composable capability. Recursive self-improvement made concrete — learn → reflect → distill into a skill → reuse → learn faster.

🦇 Batman Mode — structured (typed) cognition

  • Every memory carries a cognitive type (Observation, Reflection, Persona, Concept, Entity, Claim, Relation, Event, Conversation, Decision) plus a lifecycle state, enforced by a CI acceptance gate.
  • The system reasons differently about a raw observation vs. a derived reflection vs. an authoritative decision — and it's what makes the anti-bias machinery possible.

⚖️ Rules, standards, policy engine & governance enforcement

  • Signed substrate rules (L1–L6) requiring the operator's cryptographic key.
  • Namespace standards — per-domain policy with a fail-closed default.
  • Tri-state enforcement (off / advisory / enforce) with a mandatory-hook presence gate: under enforce, a write whose required governance hook is missing is denied. You can't accidentally run ungoverned.

🔐 Cryptographic attestation + encryption

  • Ed25519 signing of memories, links, and every audit event; a tamper-evident hash chain across rows.
  • Attestation levels — claimed vs. agent-attested vs. signed-by-peer; forged signatures rejected unconditionally.
  • Identity chains — mTLS, cert-chain delegation with namespace confinement, replay-nonce protection.
  • Encryption — at-rest (SQLCipher) + per-memory encrypted envelope (lossless archive/restore) + TLS in transit + TLS-framed syslog for off-host audit shipping.

🧬 Anti-bias & anti-echo-chamber for recursive AI the differentiator

The danger: if one LLM family does all the reflecting, its blind spots get laundered into "authoritative" knowledge and recursively reinforced — a monoculture echo chamber compounding bias into bedrock. (The DeepMind "From-AGI-to-ASI" audit's #1 structural risk.) The layered answer:

  • Decorrelation visibility — continuously measures producer dominance across the reflection corpus; raises an advisory when knowledge is dominated by one source, with an honest "claimed-not-attested" caveat so it never gives false confidence.
  • Model-family attestation — bind a reflection to a verifiable model family, so "distinct" means provably-different-model, not a forgeable label.
  • Heterogeneous evaluator panel — route judgments through genuinely different model families.
  • N≥3 model-family-distinct quorum — a bias-displaced reflection is refused as authoritative unless ≥3 attested, model-family-distinct reflectors agree. Knowledge becomes "true" by cross-model consensus, not one model's repetition.
  • Backed by contradiction detection and bitemporal validity (valid-from / valid-until) so the substrate holds disagreement and change over time instead of flattening to one voice.
echo-chamber defense:  measure dominance → attest model family → panel of distinct models → quorum refuses single-source "truth"

✨ Other wazoo

  • Federation + W-of-N quorum — knowledge replicates across machines/regions and becomes authoritative only by quorum.
  • Native knowledge graph (Apache AGE) — typed relationships, multi-hop path-finding, timeline queries.
  • Atomisation — shatters compound memories into atomic, individually-citable facts with provenance.
  • Persona-as-artifact — generate a versioned, signed persona from an entity's memory cluster.
  • Layered capture & crash-recovery (L1–L4) — operator directives and conversation turns survive process kills via idempotent, deduplicated capture and transcript replay.
  • Forensic bundle export/verify — package and cryptographically verify a slice of the audit trail for compliance and incident response.

The NHI vision, in one line

A self-improving, self-governing AI mind that reflects on its own knowledge to get smarter — while cryptographically proving what it knows, enforcing policy it can't bypass, and actively refusing to become an echo chamber.

Explore the substrate on GitHub →