Meaning Memory Request Access

What's New in Meaning Memory v3.18 (Beta)

A rollup of everything that has landed in the Meaning Memory Engine since the private beta opened: new features, improvements, and bug fixes through v3.18. Licensed self-host, still in private beta, with stable behavior by default.

By Clinton Stark • release, beta, changelog

Since the private beta opened, the Meaning Memory Engine has been shipping steadily. This post rolls up what has landed through v3.18.0, the first stable (non-release-candidate) wheel of the beta cycle.

Two ground rules hold across everything below. First, nothing changes default behavior: new capabilities ship behind flags or as optional extras, so an existing deployment stays the same until an operator opts in. Second, this is still a licensed self-host engine in private beta. You run it inside your own infrastructure, and your memory data stays yours.

The headline this release is Episodic arc weaving (opt-in beta), which connects related events across many sessions into coherent narrative threads. That one is significant enough to deserve its own write-up, so a dedicated explainer is coming next in the Meaning Memory Explained series. Below is the full release picture.

New features

  • Episodic arc weaving (opt-in beta). The consolidation pipeline can connect event-memories that share a coherent spine (a ticket, a project, or a recurring entity) into persisted arcs, and render them as an Open Threads section in the compiled memory file. Anchor mode is deterministic and needs no external LLM. Off by default; enable with MM_E_ARC_WEAVE_ENABLED=1, MM_E_ARC_MODE=anchor, MM_E_ARC_COMPILE=1.
  • Bundled STARE scorer (no API key required). A small, self-contained scoring model ships in the wheel, so STARE 5D scores (Significance, Temporal, Asymmetry, Relational, Episodic) can be computed without configuring an external LLM provider. It is a compact model we trained in-house through LoRA distillation cycles, purpose-built for STARE scoring rather than a general model behind a prompt. A bring-your-own extractor is still supported; the bundled scorer is the zero-dependency default.
  • Memory Seeding v0.1. Bulk-load standing knowledge into an agent’s memory from a seed corpus, so a new agent starts already knowing your domain instead of learning it one conversation at a time. Strong seed material includes internal policies, product documentation, support runbooks and playbooks, compliance and regulatory references, and curated FAQs. Provenance tracking keeps seeded memories distinguishable from what the agent learns on its own.
  • Read-only web admin dashboard (optional [dashboard] extra). Operator and evaluation visibility: metric tiles, a STARE 5D readiness row, a paginated memory browser with filters and a significance heatmap, and a per-memory detail view with a STARE radar chart and a significance-composition breakdown. Read-only by construction, binds to localhost by default, all assets vendored for air-gapped installs. Ships with a hardened systemd unit.
  • Persistent worker pool (opt-in). The consolidation coordinator can keep warm worker processes alive across runs instead of forking a fresh process each time, with a configurable recycle policy and crash isolation. Off by default; fork mode stays the default until you opt in via MM_COORDINATOR_EXECUTOR_MODE=pool.
  • Semantic dedup at consolidation. Newly extracted memories are checked against existing ones for paraphrase and near-duplicate content; a match adjusts the existing memory’s significance instead of writing a redundant entry. Tunable threshold, with a kill-switch back to prior behavior.
  • Typed run manifest + read-only operator surfaces. Each consolidation run writes a versioned, typed manifest, replacing the prior multi-layer environment-composition surface with a single validated contract. Secrets are never carried in the manifest, only the names of the variables holding them.

Improvements

  • Hybrid retrieval compiler. Consolidation output reuses the vector index path for faster, more consistent results on the Postgres backend.
  • Relational multi-hop traversal hardening. Graph queries now carry a per-query statement timeout and a per-depth fan-out cap, plus an opt-in truncation envelope so callers can detect and surface partial results cleanly instead of hanging on a pathological query.
  • Uniform product naming. Exception types, HTTP headers, and on-disk file conventions are now consistently named across the public surface, removing a layer of legacy-codename confusion.
  • Install and onboarding hardening. The Quick Start install path, wheel-pin currency, the integrity-verification ladder, and the built-in memory-category set (expanded to 15 values) were all reworked for a cleaner first install.
  • Agent-quota audit metadata. Per-agent quota override records now carry actor / reason / active audit columns.

Bug fixes

  • Staleness clock-skew clamp. Staleness scoring no longer mis-ages memories when the writing host and the engine host disagree on the clock.
  • Consolidation tenant-scope alignment (v3.18.0). The arc-weave phase now resolves its tenant scope through the same resolver used by the compile and retrieval phases, so woven arcs and compiled output read from the same data partition.
  • Encoding-firewall hardening. Additional write-path normalization and validation keep stored memories byte-clean across heterogeneous client encodings.
  • A run of smaller correctness fixes across the consolidation path for more stable run-over-run output.

Common questions

Is Episodic arc weaving on by default in v3.18?

No. Arc weaving is an opt-in beta, off by default. You enable it per deployment with three environment flags (MM_E_ARC_WEAVE_ENABLED, MM_E_ARC_MODE=anchor, MM_E_ARC_COMPILE). An existing deployment is unchanged until an operator turns it on.

Do I need an external LLM API key to score STARE dimensions?

No. A compact STARE scorer ships inside the wheel, so the five STARE 5D dimensions (Significance, Temporal, Asymmetry, Relational, Episodic) can be scored without configuring an external LLM provider. You can still bring your own extractor if you prefer one.

Is Meaning Memory generally available?

Not yet. Meaning Memory is in private beta under direct, licensed distribution. v3.18.0 is the first stable (non-release-candidate) wheel of the beta cycle, but it is not yet GA.

Can I run Meaning Memory in my own infrastructure?

Yes. It is a licensed self-host engine. You deploy it inside your own infrastructure (Docker Compose, Kubernetes, or bare-metal) and your memory data stays inside your environment.

New to the category? Start with What Is AI Agent Memory?. Running a multi-agent fleet? See how agents share memory across the enterprise.

Meaning Memory is in private beta, licensed and self-hosted: you run it inside your own infrastructure. If you operate multi-agent fleets at scale and want to evaluate it, request access or email [email protected].