Private Beta · v3.11.0-rc1 · Licensed self-host · Request a license

What did your AI agent know,
and when did it know it?

The audit-grade memory engine for enterprise multi-agent fleets. Human-like memory, shared across projects, customers, and organizations. Five dimensions of meaning (STARE). Deterministic compile. Every memory traces to source. You run it in your infrastructure (Docker Compose, K8s+Helm, or bare-metal). OTEL-native; adapter-friendly.

Read the docs · Request a license · Try the demo (below)

Why Audit-Grade Matters

Vector stores answer "is this memory relevant?" Meaning Memory answers "what did your agent know, when did it know it, and how did that knowledge change?" The first is a search problem. The second is a governance problem. Production AI deployments need both.

100%

Provenance on Every Memory

Every entry carries source, mem_id, run_id, and full lineage from observation to compile. Trace any line in the compiled snapshot back to the journal entry that produced it. No surface-level metadata timestamps, full audit chain.

Byte-Identical

Deterministic Compile

Same ledger + same config → same MEMORY.md, byte-for-byte. Run manifest captures every input. Replay any past run and verify the output matches. Regression-test agent memory the way you regression-test code.

5-Phase

Operator-Controllable Pipeline

Drain → snapshot → extract → compile → finalize. Each phase has explicit success/failure semantics, watchdogs, and observability surfaces. SLO-eligible by design, not a black-box recall layer.

Playbooks

Vertical Specialization

Three-layer config inheritance lets you pick a vertical (Developer / Customer Service / Legal) and override only what differs. Same engine, different significance calibration per industry. Ship in hours, not weeks of tuning.

scope_groups

Shared Memory Across Agents

When your support, research, and editorial agents work for the same customer, they share what matters via typed-scope memory groups. Per-agent significance on the same fact: Pris's "high" can be Molty's "low", same memory, different priorities. Multi-tenant SQL-layer isolation, GIN-indexed.

OTEL

Observability-Native

26 MCP tools instrumented with OpenTelemetry. Tenant-aware spans, tool-call lineage, configurable exporters. Phoenix, Datadog, Honeycomb, Tempo. The audit trail is not a separate system; it's the same telemetry your platform team already reads.

The 5D Memory Engine: STARE

Vector stores treat memory as a single dimension: relevance. Real memory has five. Meaning Memory is the only memory layer with all five as first-class primitives, grounded in cognitive-science principles, shipped at GA maturity in v3.10.

STARE is a mnemonic, the easy way to remember the 5 dimensions in pitches, demos, and partner conversations.

STARE 5D pentagon. Significance, Temporal, Asymmetry, Relational, Episodic STARE 5D meaning S Significance T Temporal A Asymmetry R Relational E Episodic

STARE 5D, each vertex is a first-class architectural primitive in the engine, not a marketing label. Schema columns, retrieval filters, scoring components.

S
Significance
The weight a specific person assigns to an experience, independent of objective importance. A birthday and a deploy command shouldn't tie in cosine similarity. Significance is first-class, with dual-sig (self / external) + audit ledger of every adjustment.
Significance Scorer
T
Temporal
How significance changes over time, not just decay, but sometimes growth. Decay curves are configurable per-category; validity windows answer "what was true on Tuesday?" Significance is signal-not-mutation, the trajectory survives.
Temporal Transformer
A
Asymmetry
Humans accumulate continuously; AI resets between sessions. The bridge preserves human-flagged significance across sessions, and applies per-agent scaling. The same memory ranks differently for an editorial agent versus an ops agent.
Asymmetry Bridge
R
Relational
Meaning that emerges from connection. Typed-edge graph links memories to agents, documents, customers, projects, commitments, queryable as "which memories link to customer X?" Co-witnessed moments amplify significance.
Relational Tagger
E
Episodic
Memory clusters into episodes, bounded experiences with a start, an end, a story arc. Episode summaries surface in retrieval; the LLM extractor mints episode IDs from natural conversation flow without requiring explicit grouping.
Episodic Classifier

Each dimension is an architectural primitive, schema column, retrieval filter, scoring component, not a marketing claim. See the engine architecture for the full per-dimension implementation map.

The Five-Phase Pipeline

0
Drain
1
Snapshot
2
Extract
3
Compile
4
Finalize

Every run = (ledger snapshot + config) → byte-identical MEMORY.md. Each phase has explicit success/failure semantics, watchdogs, and observability surfaces.

  1. Drain Pull pending observations from your agent's journal into the append-only ledger. Cheap, deterministic, runs every few minutes.
  2. Snapshot Capture the pre-consolidation state. The compile output is a function of (ledger snapshot + config), both versioned, both replayable.
  3. Extract LLM-driven fact extraction with significance scoring. Calibrated against domain-specific Playbooks. Parse failures surface as observability, not silent loss.
  4. Compile Budget-bounded MEMORY.md emitted under operator-controlled significance floor. Deterministic, replays produce byte-identical output.
  5. Finalize Watermark advances on success. Run manifest emitted with full lineage. Sentinel written. Watchdog disarmed for the next 24h.

Adapters & Stack: Sits on Top of What You Have

Meaning Memory is the engine, not the storage layer. Run with zero dependencies (FileBackend) or sit on Postgres (PostgresBackend, production default). An adapter layer integrates with the agent-memory and storage tools you already run; observability is OpenTelemetry-native and exporter-agnostic.

┌─────────────────────────────────────────────────────┐ │ Your AI Agent Fleet │ │ (Any agent framework · MCP · REST · Python) │ └─────────────────────┬───────────────────────────────┘ │ MCP · REST · Python ┌─────────────────────▼───────────────────────────────┐ │ Meaning Memory Engine │ │ ┌───────────────────────────────────────────────┐ │ │ │ STARE 5D primitives | 5-phase pipeline │ │ │ │ Audit-grade provenance | Run manifests │ │ │ │ Playbooks + decay + sig floor (operator) │ │ │ └───────────────────────────────────────────────┘ │ └────┬───────────────────┬─────────────────┬──────────┘ │ Adapters │ Backend │ OTEL ┌────▼────────┐ ┌───────▼─────────┐ ┌────▼─────────┐ │ Your │ │ PostgresBackend │ │ Your OTEL │ │ existing │ │ FileBackend │ │ exporter │ │ agent- │ │ (pgvector + │ │ of choice │ │ memory │ │ GIN-indexed) │ │ │ │ layer │ │ │ │ │ └─────────────┘ └─────────────────┘ └──────────────┘

Layered, not replacement

Already running an agent-memory layer for your customer-service fleet? Keep it. Layer Meaning Memory's audit-grade consolidation pipeline on top via the adapter API. Your storage stays. The governance layer is what's new.

Backend-agnostic engine

FileBackend (SQLite, 30-second setup) for evaluation and single-host deployments. PostgresBackend (pgvector + GIN-indexed scope_groups) for multi-tenant production. Same engine, same STARE primitives, different durability profile.

OTEL-native observability

26 MCP tools instrumented with OpenTelemetry. Tenant-aware spans. Tool-call lineage. Export anywhere your platform already reads, Phoenix, Datadog, Honeycomb, Tempo. The audit trail is the same telemetry your SRE team already trusts.

Licensed self-host, strict tenant isolation

You deploy the engine in your infrastructure (Docker Compose, K8s+Helm, or bare-metal). SQL-layer + auth-layer isolation between tenants, your data never leaves your VPC. Industry analog: Neo4j Enterprise / Elasticsearch (Elastic License) / Stardog. Optional vendor-hosted managed cloud is a Phase 2+ option; not v3.10 RC scope.

Operator-Controllable, 30 Seconds to Tuned

Significance floor, decay curves, vertical playbooks, retrieval blends, configured in plain TOML and env vars, not buried in code. Every knob has a default; every default is documented; every change is observable.

# Backend + tenancy (env) MM_BACKEND=postgres MM_POSTGRES_DSN=postgresql://... MM_AGENT=my-agent-id # Significance floor, what gets compiled (default 0.30, regulated verticals raise it) MM_SIG_FLOOR=0.50 # STARE feature flags (default-on in v3.10) MM_SEMANTIC_LANE=on # S-dim hybrid retrieval (lexical + pgvector RRF) MM_R_GRAPH_ENABLED=on # R-dim typed-edge relationship graph MM_READ_AUDIT_SAMPLE_RATE=1.0 # SOC2-ready read-path audit (LAB-688) # Playbook (TOML, pick a vertical, override what differs) # customer-kit/playbooks/v2-examples/legal.toml: [policies.mandatory_metadata] decision = ["party", "status"] # compile-time enforced contract [decay] default_curve = "exponential" half_life_days = 180 # legal: long memory

Playbooks: Vertical Calibration

Three vertical playbooks ship today: Developer, Customer Service, Legal. 3-layer config inheritance (base < playbook < agent < env). Pick a starting point, override what differs. Same engine, different significance calibration per industry.

Memory Constitution (v3.10+)

Playbooks are enforced compile-time contracts. A Legal playbook can declare "all decision entries require party and status metadata before compile." Phase 4 enforces the policy at compile time.

How Meaning Memory Compares

Vector stores and KV caches are excellent at storage and retrieval. Meaning Memory is what sits between your agents and that storage layer: the audit-grade consolidation pipeline. Different problem, complementary solution.

Capability Vector Memory Layer + Meaning Memory
Audit trail Metadata timestamps Full provenance, every fact traces to source
Reproducibility Best-effort top-K (drift over time) Byte-identical compile (manifest-replayable)
Operator semantics Black-box recall 5-phase pipeline, SLO-eligible, observable
What to keep Everything (or recency-based) Significance-weighted with operator-controlled floor
Vertical specialization Generic / one-size config Playbooks (Developer / Customer Service / Legal)
Multi-agent Shared storage (same view for all) Per-agent significance (same facts, different priorities)
Human feedback Binary remember/forget Gradient scoring that trains the calibration
Deployment Vendor-specific backend Standalone (SQLite) or layered on Postgres + pgvector / Qdrant / your vector store

Category snapshot of capability gaps Meaning Memory addresses on top of typical vector-store / KV-cache memory layers.

By the Numbers

Production telemetry from our internal agent fleet on continuous consolidation. Tracking what the engine actually does, not what we wish it did.

5D STARE primitives Significance · Temporal · Asymmetry · Relational · Episodic
3,566+ Tests green Engine, coordinator, consolidation, REST, MCP
26 MCP tools instrumented OTEL-native; Phoenix / Datadog / Honeycomb / Tempo
100% Provenance Every memory carries source + run_id + lineage
~3s Per-agent drain p50 Production-measured, full Phase 0–4 cycle, PG backend
<50 Pooled DB conns @ 30-agent load PgBouncer transaction-mode; 4× headroom validated
3 Vertical playbooks shipping Developer · Customer Service · Legal · more coming
v3.11.0-rc1 Current release STARE 5D GA · 3 framework adapters · v3.11 GA gated on license + counsel

Production telemetry as of May 18, 2026. Numbers track engine main; licensed-self-host engine distribution launching with v3.11.0-rc1 (private beta, request a license below).

Architecture at a Glance

Three-layer stack: your agent fleet → MM coordinator + MCP → STARE engine + storage. Detailed architecture diagram and API reference ship with private-beta onboarding.

Architecture diagram, coming soon Agents → Coordinator + MCP → STARE Engine → Storage (PG + pgvector, GIN-indexed scope_groups) Full diagram + adapter map shipping with private-beta API kit.

Engine Highlights

Ten capabilities that distinguish Meaning Memory from a vector store + a Postgres install.

STARE 5D scoring

Every memory scored on 5 dimensions at extract time; the dim signature follows the entry through the entire pipeline.

Deterministic compile

Same ledger + same config → byte-identical MEMORY.md. Replay any past run and verify the output matches.

5-phase pipeline

Drain → snapshot → extract → compile → finalize. Each phase has watchdog + sentinel + manifest emit.

scope_groups isolation

SQL-layer + auth-layer multi-tenant isolation. GIN-indexed for sub-ms scope queries even at fleet scale.

Live tuning

Sig floor, decay curves, retrieval blends, all configurable per tenant. Changes propagate without a restart.

Adapter-friendly

Sit on top of your existing memory layer, vector stores, KV caches, raw Postgres. Bring the storage you already trust.

Memory Constitution

Playbooks are compile-time contracts. Phase 4 enforces policy on every entry before it lands in the projection.

📊

SOC2-ready audit

Read-path audit sampling (LAB-688). Every retrieval is observable; every compile is replayable.

Quarantine + alerting

Bad journal entries isolate (not drop); structural drift fires Pushover within one drain cycle.

Grounded in cognitive science

STARE's 5D model is rooted in cognitive-science principles, designed for how meaning actually works, not just how data retrieves.

What Meaning Memory Is

A summary of what the engine provides.

  • An audit-grade memory engine for production AI agent fleets
  • The governance layer between your agents and storage
  • 5D meaning primitives. Significance, Temporal, Asymmetry, Relational, Episodic
  • A licensed self-host engine, you deploy in your VPC (Docker Compose, K8s+Helm, or bare-metal); strict tenant isolation; you keep your data
  • Operator-controllable, sig floor, decay curves, vertical playbooks, all configurable per tenant
  • Adapter-friendly, sits between your agents and your storage of choice (Postgres, pgvector, or your existing memory layer)
  • Grounded in cognitive-science principles
  • Dogfooded, internal production agent fleet on continuous consolidation since early 2026
  • Observable, every retrieval traces to source; every compile is reproducible byte-for-byte
  • Open SDK, licensed engine, customer SDKs (Python, TS) are open-source; the engine itself is proprietary and license-gated

Request a license

Private beta is invite-based, we're onboarding a small cohort of enterprise multi-agent fleets while STARE 5D calibration matures. Tell us about your fleet and we'll get back to you with a license, the customer-kit (wheel + INSTALL.md + reference deployments), and integration guidance. Licensed self-host engine (you deploy in your infrastructure) · audit-grade · OTEL-native. Industry analog: Neo4j Enterprise / Elasticsearch (Elastic License) / Stardog.

Want to kick the tires before contacting us? A sales-demo instance is available on request, single-tenant, hosted on cloud (Hetzner / GCP), synthetic data only, no SLA. Not a production environment.