v0.8.0 — trustworthy memory: durable, provable, conversational (+ OM Mail preview)

Release Notes — v0.8.0

Theme

v0.7.0 made reflection scale. v0.8.0 makes memory trustworthy: durable (backup and restore), provable (provenance, scope governance, conflict surfacing), and conversational (om talk) — plus an experimental preview of OM Mail, which gives agents email inboxes so they can exchange memory across machines, harnesses, models, and vendors.

Everything is additive. Defaults are unchanged, and OM Mail is CLI-only and off until you set it up.

Durability — om backup / om restore (Gate 1)

Your memory can now defend itself against a bad write or an accidental delete:

om backup --reason pre-experiment   # on-demand snapshot
om backup --list                    # what snapshots exist
om restore --latest                 # byte-faithful restore (safety snapshot first)

• Automatic pre-reflect snapshot. Every om reflect saves the last-good state before it writes, so a bad reflection can always be rolled back.
• Restore is byte-faithful and takes a safety snapshot of the current state first.
• Rotating retention bounds disk use. Snapshots are host-local and never synced.

Talk to your memories — om talk (#80, Gate 2)

A spoken-style conversation with your own memory. Each turn runs recall in the background and grounds the reply in what it finds:

om talk --query "what was I working on last week?"

v0.8.0 also makes the recall layer honest under failure: "recall timed out / backend unavailable" is now distinguished from "no relevant memory found" in every recall consumer, with a OM_TALK_RECALL_TIMEOUT knob. A new Moss backend joins bm25/qmd for recall. Text-only for now; voice providers are planned on the same loop.

Provenance and scope governance (Gates 3–4)

Memory you can audit, with sharing rules that fail closed:

• Section provenance stamps. Each reflection section carries last_reflected and derived_from_obs_window — when it was last derived and from which observation window. Retrieval objects carry typed owner/scope/source copies; inline Markdown stays authoritative.
• Pluggable scope governance. All share-out paths — cluster snapshots, cloud search upload, and OM Mail — route through one resolver with a single allowlist (SHAREABLE_SCOPES). Unknown scopes are denied by default; scope=local content never leaves your machine by any path.

Conflict surfacing — om reflect --check-conflicts (Gate 5)

Reflection can silently change a high-stakes fact. Now you can see it:

om reflect --check-conflicts            # reflect as normal, plus a conflict report
om reflect --dry-run --check-conflicts  # audit only — nothing is written

The check itself is a read-only advisory that always exits 0: it diffs the prior reflections.md against the new one and flags explicit-id divergence, slot divergence, and guardrail downgrades, writing a throwaway review artifact. Note that without --dry-run the reflection still writes as usual — the flag adds the report, it does not make om reflect read-only. Precision-first: it flags only what it is confident about. --json for machines.

Growth measurement in om doctor (Gate 6)

om doctor and om context --quality-report now measure memory growth: per-document and per-section sizes, the largest section's share, and how cold each section is — from real timestamps only, never a guess. Read-only and fail-soft; it exists so any future compaction decision is grounded in data instead of speculation.

Experimental preview — OM Mail (#88)

Agents get their own email inboxes and exchange memory as structured, signed, encrypted messages — notes, context packs, and recall requests ("ask my peer's memory a question"). Email makes the exchange durable, globally addressable, and vendor-neutral: it works across machines, harnesses (Claude Code, Codex, Grok, Cowork, Hermes), models, and orgs with no shared infrastructure beyond a mailbox.

Trust model in five bullets:

• The mail provider is an untrusted carrier; inbox access is not trust.
• Every envelope is Ed25519-signed and verified against a peer key you pinned locally.
• Context packs are always encrypted with a key exchanged out of band.
• Fail closed: unknown sender, bad signature, or undecryptable payload is held for review, never ingested.
• Nothing auto-executes, and scope=local never leaves the host — including through recall answers.

Try it with two agents on one machine (no account needed, localdir provider), or live with AgentMail inboxes: see docs/mail-memory.md.

v0.8.0 also ships the public plugin seams: out-of-tree mail providers and CLI add-ons plug in through entry points (observational_memory.mail_providers, observational_memory.cli_plugins), with contributor terms in CONTRIBUTING.md. Pairwise mail is MIT core and stays that way; team-fabric features (digests, trust roots, group addresses) are planned as separately licensed add-ons on these seams.

Smaller fixes

• om reflect --async now rejects a cross-provider reflector model (e.g. a leftover grok-* or Claude model) before submitting the OpenAI Batch job, instead of letting the batch fail hours later (#77).
• The test suite is hermetic by default: ambient OM_* and provider env vars no longer leak into test runs.

Compatibility

• All changes are additive; no defaults changed. New surfaces (om backup, om talk, om mail, --check-conflicts) activate only when you call them.
• One new background behavior: om reflect takes an automatic pre-reflect snapshot. Retention keeps disk use bounded.
• OM Mail is experimental: the CLI surface may evolve, and it is CLI-only (no hooks, no daemon, polling only).

Validation

• Full suite green on Python 3.11/3.12/3.13 in CI; 1,059 tests at release.
• Every feature PR passed an adversarial review pass (Codex) before merge; findings (e.g. an AgentMail pagination-order bug, a timestamp-tie sync bug, a localdir path-escape, a fine-tuned-model guard gap) were fixed with regression tests.
• OM Mail was validated live across two machines (laptop ↔ Linux VM over real AgentMail inboxes): note flow with held→accept→provenance, recall request/response, encrypted context pack, and a wire-level check that only ciphertext crosses the provider.

Out of scope / next

• v0.9.0 is reserved for OM Mail GA — handshake tokens, live listening, digests: see plans/om-mail-next-steps.md.
• The recall-quality benchmark (#81) runs in observational-memory-bench and will inform a future decision on default backend / fusion.
• The full memory-unit store and destructive compaction remain deferred behind Gate-6 data (#71).