v0.2.5 ships Mnemon Memory and the local Agency authority in one mnemon binary.

Changelog

• eff9a46 feat: add MiniMax Code memory integration
• b64fec4 feat: add ZCode memory integration
• 0fcb3b6 fix(setup): support ZCode hooks on Windows

v0.2.4 ships Mnemon Memory and the local Agency authority in one mnemon binary.

Changelog

• f30d72b feat(memory): make the auto-prune insight ceiling configurable
• f89f950 fix(artifact): report creation when a Put promotes another writer's marker
• 515909b fix: resolve latest dsh-mnemon release

v0.2.3 ships Mnemon Memory and the local Agency authority in one mnemon binary.

Changelog

• 4b6da53 fix: publish Windows release artifacts

v0.2.2 ships Mnemon Memory and the local Agency authority in one mnemon binary.

Changelog

• 1811c88 feat: add direct DSH plugin install
• 25f098e fix: disable Agency on Windows

v0.2.1 — Memory Store Reliability

v0.2.1 strengthens Mnemon Memory's SQLite maintenance paths without changing the CLI or public command structure.

Memory reliability

• Auto-prune now records one prune oplog entry for every soft-deleted insight, including the active count and configured capacity.
• Repairs the legacy edges table migration so schema detection works with foreign-key enforcement enabled.
• Allows the edge-table rebuild to complete when recovered stores contain orphaned edges, preserving those edges verbatim instead of silently deleting them.
• Makes the migration probe rollback-only and collision-resistant, so failed migrations cannot leave probe rows behind or remove real edges whose IDs resemble the old sentinel.

Existing stores already using the current four-edge schema are unchanged. For legacy stores whose constraint still permits the obsolete narrative edge type, the migration applies the existing four-graph policy: narrative edges are removed while all other edges are retained.

Internal maintenance

• Groups Memory and Agency implementation packages under explicit domain boundaries. This structural change does not alter CLI commands, wire protocols, state paths, or host integrations.

Release artifacts are available for macOS and Linux on amd64 and arm64.

Install

brew install --cask mnemon-dev/tap/mnemon
# or
go install github.com/mnemon-dev/[email protected]

Full changelog: v0.2.0...v0.2.1

中文摘要

v0.2.1 聚焦 Memory SQLite 存储可靠性：自动裁剪现在会逐条写入 oplog；旧版 edges 表迁移可在启用外键及存在孤立边的情况下安全完成，并避免探针残留或误删真实数据。本版本同时整理了 Memory / Agency 的内部包边界，不改变 CLI、协议或状态路径。

v0.2.0 — Mnemon Agency Preview

v0.2.0 integrates Mnemon Agency Preview into the single mnemon executable while keeping Mnemon Memory as the established root command surface.

Mnemon Agency Preview

• Adds project-local durable responsibility and admitted effects for existing Agent Runtimes.
• Introduces the bounded View → Intent → Receipt interaction loop.
• Persists three small collaboration objects: Handling, Artifact, and Reference.
• Adds a Pi-first integration through project-local hooks and guidance.
• Adds explicit, authenticated, pairwise peer exchange. Remote contributions remain candidates and must be admitted by the receiving node.
• Exposes setup, daemon lifecycle, and peer configuration through mnemon agency ....

mnemond is the local daemon and protocol role behind Agency. It is not a second executable, and it does not replace an Agent's planner, tools, provider, or model runtime.

Memory remains the mainline

Agency is opt-in and remains separate from Memory:

• existing Memory commands stay at the mnemon root, including remember, recall, and setup;
• named stores, the Memory SQLite data model, and existing Memory setup integrations remain unchanged by Agency;
• mnemon setup continues to configure Memory;
• mnemon agency setup separately enables Agency for one project;
• Agency state is stored separately under the project's .mnemon/agency directory.

This release also fixes a Memory safety issue: conflict candidates are no longer silently removed, and cosine-only similarity is no longer sufficient to replace an existing memory.

Preview and compatibility notice

Agency is actively evolving and is intended for evaluation and controlled use. Its command surface, Runtime integration, protocol, peer exchange, and on-disk/wire formats may change before an explicit stability milestone. This Preview status applies to Agency; Memory retains its separate product contract.

The obsolete experimental Harness surfaces and event-emission bridge, including mnemon event and MNEMON_HARNESS_EVENT_EMIT, have been retired without compatibility shims.

Release artifacts are available for macOS and Linux on amd64 and arm64. Windows archives are not published for v0.2.0.

Install

brew install --cask mnemon-dev/tap/mnemon
# or
go install github.com/mnemon-dev/[email protected]

Existing Homebrew Formula installations are migrated to the v0.2.0 Cask through the tap migration record.

Validation

• deterministic CI and Go test suite;
• native Memory CLI E2E: 151/151 checks;
• Agency race, process, Docker, peer, and Pi Runtime boundaries;
• real multi-service, multi-network domain-operations scenario.

Full changelog: v0.1.17...v0.2.0

中文摘要

v0.2.0 首次把 Mnemon Agency Preview 集成到单一 mnemon 可执行文件中。Memory 的根级命令、Store、SQLite 数据和既有 setup 链路保持独立；Agency 通过 mnemon agency ... 按项目单独启用，状态保存在 .mnemon/agency。

Agency 引入 View → Intent → Receipt 循环、Handling / Artifact / Reference 三类协作对象、Pi-first Hook 接入和显式 Peer 协作。Agency 仍处于持续迭代的 Preview 阶段，其命令、Runtime 接口、协议及持久化/传输格式可能继续变化；这不改变 Mnemon Memory 的独立产品契约和主使用链路。

v0.1.17 broadens setup support for additional coding hosts, fixes hook/runtime edge cases, and advances the experimental harness runtime work.

Note: harness modules, harness documentation, and harness evaluation assets in this repository remain experimental and are not part of this release's public stability guarantee.

Highlights

• Added setup integrations for Kimi Code and OpenCode, expanding the supported host surfaces installed by mnemon setup.
• Fixed Claude memory subagent guidance so delegated memory writers run mnemon remember / mnemon link directly instead of recursively spawning another subagent.
• Fixed Hermes hook IO decoding so UTF-8 content is handled correctly.
• Improved user-facing CLI behavior around service help, daemon readiness checks, occupied listen addresses, and hidden legacy/debug harness commands.

Experimental harness work

• Added and refined Multica runtime adapter flows, participant provisioning, managed wake handling, progress projection, and R3 provider-surface gating.
• Added Cloudflare MnemonHub bootstrap and local backend support behind the experimental harness layer.
• Added GitHub remote workspace and publication-store infrastructure for multi-agent mesh validation.
• Reworked event-system and presentation boundaries around mnemond, mnemonhub, hostagent, runtime surfaces, and derived event views.
• Expanded acceptance, Docker, live, ABI, and boundary tests around Multica, GitHub mesh, managed runtime, and public command surfaces.

Fixes

• Preserved setup hook executable bits and hardened hook execution paths.
• Stabilized GitHub publication listing, polling, sync cursors, rate-limit handling, and acceptance shutdown paths.
• Hardened Multica routing metadata, assignment dispatch, feedback target recovery, overlap validation, projection text, and hub-flow prerequisites.
• Redacted bearer tokens from activation traces and rejected invalid daemon role/config states.

Validation

• CI passed for the final merged PR included in this release.
• Local validation before tagging: go build -o mnemon ..

Merged PRs

• #66 feat(setup): add Kimi Code integration
• #67 feat(setup): add OpenCode integration
• #68 feat: productize AgentTeam-oriented harness structure
• #69 feat: add realistic project validation
• #70 feat(harness): add GitHub remote workspace backend
• #71 feat(harness): redesign R2 event interactions
• #72 feat(harness): add managed runtime readiness
• #73 feat: add Multica runtime adapter
• #74 feat: clean up R2 event-system CLI cluster
• #75 fix: harden Multica overlap validation
• #76 refactor: reshape R3 Multica runtime surface
• #78 feat: land R3-1 Cloudflare hub and Multica runtime gate
• #80 fix(hermes): decode hook IO as UTF-8
• #81 fix(setup): keep Claude memory subagents as leaf writers

Full changelog: v0.1.16...v0.1.17

v0.1.16 focuses on stable CLI, storage, and entity-graph behavior.

Note: harness modules, harness documentation, and harness evaluation assets in this repository remain experimental and are not part of this release's public stability guarantee.

Changelog

• 23ec779 docs(readme): link Claude Code section
• aa3fba9 docs(readme): mention TRAE Work setup support
• 2d50be8 feat(setup): add CodeBuddy integration (#64)
• 8731b3f feat(setup): add Qoder and QoderWork integrations (#62)
• 4187814 feat(setup): add Trae integration
• 50f9e49 feat(setup): add WorkBuddy integration (#65)

v0.1.15 focuses on stable CLI, storage, and entity-graph behavior.

Note: harness modules, harness documentation, and harness evaluation assets in this repository remain experimental and are not part of this release's public stability guarantee.

Changelog

• 0a7a7da Potential fix for code scanning alert no. 1: Workflow does not contain permissions
• b375cfa docs(core/kernel): scope DeferralCount's equivalence note to the reconciler path
• 72779c0 docs(core/standard): clarify the adapter OpID is illustrative, not a per-proposal uniqueness key
• c6eeea0 feat(core/callback): registry + builtin callbacks; intent-not-fact; Dispatch drops erroring callbacks
• 9a5b2e0 feat(core/callback): trusted-author subprocess callback (stdin event -> stdout intents); fail-closed
• 5b08e9c feat(core/contract): pin control-plane wire types
• 66251c3 feat(core/kernel): Apply = atomic multi-resource CAS + decision-status state machine
• fc40265 feat(core/kernel): schema guard + authority enforce + sentinel errors
• f71df0e feat(core/kernel): single-conn sqlite store with per-resource CAS + persisted event/decision logs
• 8b572bd feat(core/projection): generated view with version digest + pull-feedback channel
• 306758e feat(core/reconcile): ResolveModes selects from the fixed catalog (define≠select); scripts rejected
• f5e655f feat(core/reconcile): deterministic RunOnce + trusted opFromEvent + liveness escalation; conflict arms A-E
• 36594df feat(core/standard): contract-only second adapter + falsifiable smallness dep-check
• 21823d0 feat(setup): add Cursor IDE integration
• 66fae9b feat: add Local Mnemon command
• e9c2e46 feat: add local memory admission loop
• 4c3c13a feat: add remote sync API
• 12862c9 feat: add sync push command
• 129c09f feat: admit local skill declarations
• 59d3e08 feat: harden setup local integration
• c1d4d14 feat: import remote memory commits
• cd54f50 feat: record pending sync commits
• bedccd1 feat: run local memory admission through governed wasm
• 23fb1d2 feat: run local skill admission through governed wasm
• a1092be fix(core): empty CorrelationID must not share one escalation bucket (R2#1)
• e05ff36 fix(core): escalation counts only rebase deferrals; migrate added decision columns (verification round 2)
• a12249e fix(core): harden malformed-op rejection + lock the feedback tiebreak (verification LOW findings)
• 66a39bc fix(core): reject empty/malformed ops instead of rubber-stamping an Accepted no-op
• 5d537a9 fix(core): stamp Decision.IngestSeq from the triggering event (audit link + stable feedback order)
• c7a216c fix(core/contract): shrink AuthzCatalog to the implemented mode only (define=select honesty)
• baab978 fix(core/kernel): fail on event marshal/decode errors instead of corrupting the ingest stream
• 5f16105 fix(core/reconcile): derive liveness-escalation count from the durable log (Invariant #10 survives restart)
• 1ab5479 fix(core/reconcile): durable cursor seeded from the decision log (no restart replay)
• f8c8846 fix(core/reconcile): escalation requires a declared CorrelationID (replace ev.ID fallback)
• 82804d7 fix(core/reconcile): reconcile only proposal events; observations are not write attempts
• 89309c0 fix: hide internal setup vocabulary
• 60c021d fix: keep cli core access behind server
• 52539b6 test(core/reconcile): conflict mode selects distinct deterministic decisions; auto_merge_disjoint fail-closed

Added

• mnemon setup --target cursor now installs a Cursor integration, including a Mnemon skill, prompt files, and Cursor lifecycle hooks under .cursor/ or ~/.cursor/. The session-start hook primes Cursor with Mnemon guidance and memory status, the stop hook can nudge for durable-memory writeback, and an optional pre-compaction hook can remind users to preserve durable context.
• mnemon setup --eject --target cursor removes the installed Cursor skill and Mnemon hook entries while preserving unrelated Cursor hook configuration.
• Cursor environment detection is included in interactive setup when the cursor binary or ~/.cursor/ config directory is present.

Fixed

• CI workflow permissions are now declared explicitly for the code-scanning workflow surface.

Tests

• Added Cursor setup coverage for skill installation, executable hook writes, hook registration, preservation of unrelated Cursor hooks, and eject cleanup.

Note

• Harness modules remain experimental and are not part of the v0.1.15 public stability guarantee. This release includes substantial internal harness iteration, documentation, and validation work, but release-path commands such as mnemon setup continue to keep formal integrations under cmd/ and internal/.

v0.1.14 focuses on stable CLI, storage, and entity-graph behavior.

Note: harness modules, harness documentation, and harness evaluation assets in this repository remain experimental and are not part of this release's public stability guarantee.

Highlights

• Known-entity extraction now has an index-aware path that improves recall for seeded project names, internal codenames, and single-segment CamelCase entities without globally loosening first-mention extraction.
• Embedding blobs are stored as little-endian float32 values going forward, reducing embedding storage size by roughly half while preserving float64 vector APIs and cosine calculations.
• Claude Code setup now handles the $HOME/.claude collision case by writing absolute hook commands when an apparent project-local install is actually the user-global Claude config.

Added

• Entity extraction now admits wider candidate tokens when they already exist in Mnemon's known-entity index. This lets a name seeded through provided entities or earlier extraction propagate through later insights and queries without treating arbitrary new wide tokens as first-class entities.
• Store support for loading known entities from active insights. The graph engine uses this index to seed the new extraction path while continuing to ignore soft-deleted records.

Changed

• New embeddings are serialized as float32 blobs instead of float64 blobs. For common 768-dimensional embeddings, this lowers each stored embedding from 6144 bytes to 3072 bytes.
• Existing official databases migrate legacy float64 embedding blobs to float32 on open. The migration records completion in SQLite PRAGMA user_version and leaves malformed or non-legacy-looking blobs untouched rather than blocking startup.

Fixed

• Running mnemon setup for Claude Code from $HOME no longer writes relative hook commands into the user-global ~/.claude/settings.json. Mnemon now detects this collision, resolves symlinks, honors CLAUDE_CONFIG_DIR, and writes absolute hook paths so hooks work from any future session directory.
• Genuine project-local Claude Code installs continue to use relative hook commands, preserving the existing local-project behavior.

Tests and Validation

• Added vector serialization tests for float32 round trips, invalid blob lengths, and legacy float64 deserialization.
• Added store migration tests for normal legacy embeddings, malformed blobs, and non-legacy-looking blobs that should be skipped.
• Added entity extraction and store tests for known-entity propagation and soft-delete filtering.
• Release validation completed successfully:
  • go build -o mnemon .
  • make test (147 E2E checks)
  • go test ./... in a clean worktree
  • GitHub Release workflow for v0.1.14
  • GitHub master CI for commit 52fdd88

Artifacts

• macOS: mnemon_0.1.14_darwin_amd64.tar.gz, mnemon_0.1.14_darwin_arm64.tar.gz
• Linux: mnemon_0.1.14_linux_amd64.tar.gz, mnemon_0.1.14_linux_arm64.tar.gz
• Windows: mnemon_0.1.14_windows_amd64.zip
• Checksums: checksums.txt
• Homebrew tap updated via GoReleaser.

Changelog

• 9edfc4a Switch to float32 for storage (#55)
• 545f625 feat(graph): add index-aware fourth path to entity extraction
• 1982cff fix(setup): guard against project-local install colliding with the user-global Claude config (#53)

Added

• Entity extraction now has an index-aware fourth path for known entities. When Mnemon has already seen an entity through provided metadata or earlier extraction, later graph operations can admit wider candidate tokens that match the known-entity index. This improves recall for personal project names, internal codenames, and single-segment CamelCase entities without globally loosening first-mention extraction.
• Store support for loading known entities from active insights, used by the graph engine to seed index-aware extraction while still skipping soft-deleted records.

Changed

• Embedding blobs are now stored as little-endian float32 values instead of float64, cutting embedding storage size roughly in half while keeping the public vector API and cosine-similarity calculations in float64.
• Existing official databases are migrated on open from legacy float64 embedding blobs to the new float32 storage format. The migration records its completion in SQLite PRAGMA user_version and leaves malformed or non-legacy-looking blobs untouched rather than blocking database startup.

Fixed

• Claude Code project-local setup now detects the degenerate $HOME/.claude collision. When mnemon setup is run from $HOME, the apparent local .claude/ directory is actually Claude Code's user-global config directory; Mnemon now writes absolute hook commands in that case so hooks resolve correctly from every future Claude Code session directory.
• Added symlink-aware and CLAUDE_CONFIG_DIR-aware coverage for Claude Code config collision detection, plus regression tests that genuine project-local installs keep relative hook paths.

Tests

• Added round-trip and invalid-length coverage for float32 vector serialization, plus legacy float64 deserialization coverage for migration.
• Added store migration coverage for normal legacy embeddings, malformed blobs, and non-legacy-looking blobs that should be skipped.
• Added entity extraction and store tests for the known-entity path, including propagation of previously seeded project names without admitting arbitrary new wide tokens.

Added

• English/Chinese import guide coverage and setup skill guidance for using mnemon import on historical chat exports.

Added

• mnemon import <file> bulk-imports memory draft JSON files into a selected Mnemon store. Imported insights use the normal write path for duplicate and conflict detection, embeddings when Ollama is available, entity/causal/ semantic graph construction, lifecycle scoring, and capacity pruning.
• internal/importdraft defines the public draft schema for historical chat imports:
  • top-level schema_version, source, insights, and optional edges
  • insight fields for content, category, importance, tags, entities, source, and RFC 3339 created_at
  • explicit edge fields for source_index, target_index, edge_type, weight, and reason
• mnemon import --dry-run validates draft structure without writing to the database, and mnemon import --no-diff inserts all draft insights without duplicate/conflict detection.
• Import documentation with the full schema reference, category and importance guidance, edge-type guidance, examples, operational notes, and a copy-paste LLM prompt for turning chat exports into memory_draft.json.

Fixed

• Historical imports no longer create incorrect temporal backbone edges when created_at is older than existing store contents. The import path disables real-time temporal edge generation, then repairs affected source timelines after all backdated insights are written so imported memories are inserted into the global chronological chain.
• Explicit draft edges now participate in effective-importance calculation before AutoPrune runs. Import finalization inserts explicit edges, repairs temporal edges, refreshes touched insight scores, and only then applies pruning.

Changed

• The graph engine now supports internal EngineOptions with TemporalMode, allowing import to skip real-time temporal generation without changing the behavior of mnemon remember.
• Store internals gained helpers for source-ordered active insight scans and typed single-edge deletion, used by temporal repair during imports.

Tests

• Added validation coverage for the import draft schema, including required content, category and importance constraints, RFC 3339 timestamps, edge index bounds, and source fallback behavior.
• Added command integration coverage for backdated import temporal repair and effective-importance refresh after explicit draft edges.

Changed

• BREAKING: recall now emits a compact, LLM-friendly JSON shape by default. The default result entries keep id, content, category, importance, intent, matched_via, confidence, and rounded score, while omitting verbose debug fields such as signals, timestamps, access_count, tags, entities, source, and traversal metadata. (#3)
• The previous full recall payload remains available with --verbose for scripts, debugging, and callers that still need meta, signals, or the full embedded insight object.

Added

• recall --verbose flag to restore the previous full recall response.

Added

• --embed-model global flag exposes the existing MNEMON_EMBED_MODEL env var on the CLI, applied consistently across embed, recall, and remember. Useful for switching to multilingual models such as nomic-embed-text-v2-moe:latest on code-switched corpora without exporting the env var. Note: switching to a model with a different output dimension silently invalidates existing embeddings — backfill after changing.

Added

• Privacy-safe memory receipts via mnemon receipt, allowing users to show memory activity with identifiers, timestamps, summaries, and relation metadata without exposing stored memory content.

Fixed

• Read-only store operations now skip access-count and oplog writes, preventing recall/search/status-style paths from attempting mutations when the store is opened in read-only mode.

Changed

• Project licensing switched to Apache-2.0.

Added

• Codex integration: mnemon setup --target codex deploys the mnemon skill, prompt files, and Codex lifecycle hooks (SessionStart, UserPromptSubmit, Stop) into .codex/ or ~/.codex/. mnemon setup --eject --target codex removes the installed Codex surface while preserving unrelated hooks.
• Nanobot integration: mnemon setup --target nanobot deploys a skill file to .nanobot/skills/mnemon/SKILL.md (local) or ~/.nanobot/workspace/skills/mnemon/SKILL.md (global, recommended). mnemon setup --eject removes it. Detection is automatic when the nanobot binary or ~/.nanobot/workspace/ directory is present.

Note

• Harness modules remain experimental and are not part of the v0.1.6 public stability guarantee. Release-path integrations are implemented under cmd/ and internal/.

Fixed

• Diff now sorts matches by Similarity descending before selecting the overall suggestion. Previously, KeywordSearch ordered candidates by token overlap score, so a high-keyword-score candidate classified as ADD could mask a lower-keyword-score candidate with higher Jaccard similarity that should have been UPDATE or DUPLICATE.
• Deduplication false positives on scientific and domain-specific text:
  • Removed bare "not" from negation words — it appears in virtually all scientific prose and caused unrelated records to be classified as CONFLICT.
  • Gated negation-word check behind similarity ≥ 0.7 — at borderline similarity, shared domain vocabulary is not a reliable conflict signal.
  • Raised cosine dedup threshold from 0.70 to 0.85 — same-domain different-fact pairs (e.g. survey records at different locations) produce cosine ~0.75 with nomic-embed-text and were incorrectly triggering UPDATE.
  • Switched token dedup from bidirectional-max (ContentSimilarity) to Jaccard (|A∩B|/|A∪B|) — penalises texts that share vocabulary but differ in most tokens, preventing formulaic records from scoring as UPDATE.

Added

• remember --entity-mode to choose merge, provided, or auto entity handling.

Fixed

• Store name validation for resolved store paths.
• Query limit validation for recall and related command paths.
• Sidecar migration error reporting.
• Setup integration cleanup reliability.

Note

• Harness modules, harness documentation, and harness evaluation assets remain experimental and are not part of the v0.1.4 public stability guarantee.

Added

• MNEMON_EMBED_DIMENSIONS env var for Matryoshka dimension truncation (e.g., 256-dim instead of 768)

Added

• OpenClaw full integration: internal hook (agent:bootstrap) + plugin (before_prompt_build)
• mnemon setup --target openclaw now auto-deploys skill, hook, plugin, and config
• Optional hook selection for OpenClaw (remind, nudge, compact) matching Claude Code parity
• Plugin version patching from binary ldflags
• LLM-supervised tagline in README

Changed

• README restructured: technical differentiator (comparison table) moved above value proposition
• OpenClaw setup no longer requires manual plugin configuration

Initial public release.

Added

• Core CRUD: remember, recall, forget, search, status, log
• Four-graph architecture: temporal, entity, causal, semantic edges
• Intent-aware smart recall with beam search graph traversal
• Built-in deduplication and conflict resolution
• Retention lifecycle: importance decay, access-count boosting, garbage collection
• Named memory stores for data isolation (mnemon store list|create|set|remove)
• MNEMON_STORE environment variable and --store CLI flag for store selection
• Automatic migration of legacy ~/.mnemon/mnemon.db to ~/.mnemon/data/default/
• Optional embedding support via Ollama (nomic-embed-text)
• Knowledge graph visualization (mnemon viz)
• Claude Code integration via hooks (prime, remind, nudge, compact)
• OpenClaw integration via skill deployment
• mnemon setup interactive installer with --eject support
• Release pipeline: GoReleaser, GitHub Actions, Homebrew tap
• Comprehensive documentation with Chinese translations