v0.3.0 — The Chronicle

Changelog

All notable changes to this project are documented in this file.
Format loosely follows Keep a Changelog.

[0.3.0] — 2026-07-12 «The Chronicle»

Added

• Session capture (#1) — samemind capture --engine <id> [--source <path>] [--since <ts>] [--dry-run] (tools/capture.mjs, docs/session-capture.md): read-only adapter framework
  that pulls a live engine's own session store into inbox/<engine>.md, closing the last
  bespoke per-engine sync bridge from dogfooding. MVP adapters: claude-code (distills each
  JSONL transcript's final assistant text + session id/project/message-count meta) and
  generic-markdown (any directory of .md diaries → title + first lines + path pointer
  notes, e.g. OpenClaw's memory/*.md). Idempotent via .samemind-capture-state.json in the
  bundle root; secret shapes (npm_/sk-/ghp_/AKIA) masked before writing; distilled text
  runs through the same injection-quarantine as memory_write_inbox; --dry-run writes
  nothing. Adding an engine is one more ADAPTERS entry.
• Event ledger (#3) — samemind ledger append|status|read (tools/ledger.mjs,
  tools/lib/ledger.mjs, docs/event-ledger.md): an append-only, fine-grained event log
  (ledger/events.jsonl) complementing the coarse work-discipline layer where Task.status
  is edited in place. append --actor <id> --topic <t> --phase start|step|done|fail|block|note [--status ok|wip|partial|fail] --action "..." [--artifact <a>] [--ref <r>] validates both
  dictionaries (rejects, never silently coerces); status surfaces 🔥 open failures — the last
  fail/block event of a topic not yet closed by a later done or status: ok event — before
  every topic's current stage; read --topic <t> prints one topic's full history. MCP gains
  memory_ledger_append/memory_ledger_status (same SAMEMIND_AGENT-as-actor and
  injection-quarantine contract as memory_write_inbox). samemind board gains a
  🔥 Open failures section above 🔴 Blocked (capped at 5, freshest first, full count in the
  heading), in both markdown and --html. ledger/ is a reserved tier like inbox//secret//
  mirror/ — never walked as graph concepts, so query validate/list/get stay unaffected.

[0.2.1] — 2026-07-12

Added

• Exclude-by-source (anti-echo, #2) — an engine no longer gets back what it just wrote.
  MCP memory_search accepts exclude_source (validated to [a-z0-9-]); recall/gde/brief
  gain --exclude-source <id>. Concepts whose frontmatter source matches the id are filtered
  from the result (works for both string and list source, in BM25 and semantic paths).
• Smooth brief budget — brief --budget no longer drops whole sections in a step curve.
  After tier selection, the last kept tier-1/2 section is trimmed by paragraphs to land within
  ±10% of the budget, marked …truncated. Tier-0 (boundaries / owner rules / engine role) is
  never trimmed. Size now grows monotonically with the budget instead of jumping.
• Generic install — install --agent <any-id> --file <path> installs into any instruction
  file for an unsupported agent (generic brief + protocol block, idempotent between the markers).
  --file is required for an unknown id; --list advertises + any id via --file.

[0.2.0] — 2026-07-12 «The Flywheel»

Added

• Knowledge-cycle layer — three new concept types close the loop from facts to plans
  (docs/knowledge-cycle.md): Analysis (facts → pattern → implications), Research
  (question → findings → verdict, with source citations), Idea with a maturity status
  (spark → incubating → adopted / rejected, rejected_reason required). Edge conventions
  over existing relations: informs (Analysis/Research → Idea), spawned_by
  (Research → Analysis), led_to (Idea → Plan).
• Ideas on the board — samemind board gains a 💡 Ideas section: incubating first,
  then sparks; adopted collapse into an "Adopted → Plans" line via led_to; rejected hidden.
• Agent reflection protocol — memory-protocol and all three snippets teach agents to
  write reflections on immature Ideas through their own inbox (target: <idea path>),
  never editing the idea file directly; curation merges into ## Reflections.
• HTML projections — samemind board --html [--out <file>] and
  samemind handoff --html: self-contained pages (inline CSS, light/dark via
  prefers-color-scheme, zero JS, zero external resources) with code-generated SVG
  visualizations (kanban bars, ideas strip, decisions timeline). Markdown stays the canon;
  HTML is always a generated projection. Board/handoff internals split into
  model + renderers so both outputs share one data path.
• Validator warnings for the new types (Idea without status, rejected without
  rejected_reason); scaffold templates for all three types in samemind init;
  demo bundle gains a linked Analysis → Research → Idea working example.

[0.1.2] — 2026-07-12

Fixed

• inbox/ is now a proper reserved tier in walk()/load() (tools/lib/okf.mjs), on the same
  footing as secret//mirror/: excluded by default, opt-in via includeInbox /
  --include-inbox (okf-query, okf-recall). Before this fix, the first ever write through
  memory_write_inbox (MCP) — whose frontmatter carries only okf_version, no type — made
  samemind query validate permanently non-conformant, because raw inbox notes were treated as
  ordinary graph concepts. (#4)
  • validate/list/links/rel/get no longer see inbox/ by default.
  • tools/consolidate.mjs keeps reading inbox/ (opts in explicitly — that's its whole
    purpose: mapping raw notes to canon gaps).
  • MCP memory_search/memory_get/memory_list never returned inbox content and still don't.
  • Added a regression test: a fresh bundle → memory_write_inbox → validate stays conformant.

[0.1.1]

Fixed

• npx/.bin symlink: resolve argv[1] through realpath so the CLI's isMain check
  recognizes itself when invoked via a symlinked bin (e.g. npx samemind).

[0.1.0]

• Initial release.

v0.2.1 — anti-echo, smooth budget, generic install

Changelog

All notable changes to this project are documented in this file.
Format loosely follows Keep a Changelog.

[0.2.1] — 2026-07-12

Added

• Exclude-by-source (anti-echo, #2) — an engine no longer gets back what it just wrote.
  MCP memory_search accepts exclude_source (validated to [a-z0-9-]); recall/gde/brief
  gain --exclude-source <id>. Concepts whose frontmatter source matches the id are filtered
  from the result (works for both string and list source, in BM25 and semantic paths).
• Smooth brief budget — brief --budget no longer drops whole sections in a step curve.
  After tier selection, the last kept tier-1/2 section is trimmed by paragraphs to land within
  ±10% of the budget, marked …truncated. Tier-0 (boundaries / owner rules / engine role) is
  never trimmed. Size now grows monotonically with the budget instead of jumping.
• Generic install — install --agent <any-id> --file <path> installs into any instruction
  file for an unsupported agent (generic brief + protocol block, idempotent between the markers).
  --file is required for an unknown id; --list advertises + any id via --file.

[0.2.0] — 2026-07-12 «The Flywheel»

Added

• Knowledge-cycle layer — three new concept types close the loop from facts to plans
  (docs/knowledge-cycle.md): Analysis (facts → pattern → implications), Research
  (question → findings → verdict, with source citations), Idea with a maturity status
  (spark → incubating → adopted / rejected, rejected_reason required). Edge conventions
  over existing relations: informs (Analysis/Research → Idea), spawned_by
  (Research → Analysis), led_to (Idea → Plan).
• Ideas on the board — samemind board gains a 💡 Ideas section: incubating first,
  then sparks; adopted collapse into an "Adopted → Plans" line via led_to; rejected hidden.
• Agent reflection protocol — memory-protocol and all three snippets teach agents to
  write reflections on immature Ideas through their own inbox (target: <idea path>),
  never editing the idea file directly; curation merges into ## Reflections.
• HTML projections — samemind board --html [--out <file>] and
  samemind handoff --html: self-contained pages (inline CSS, light/dark via
  prefers-color-scheme, zero JS, zero external resources) with code-generated SVG
  visualizations (kanban bars, ideas strip, decisions timeline). Markdown stays the canon;
  HTML is always a generated projection. Board/handoff internals split into
  model + renderers so both outputs share one data path.
• Validator warnings for the new types (Idea without status, rejected without
  rejected_reason); scaffold templates for all three types in samemind init;
  demo bundle gains a linked Analysis → Research → Idea working example.

[0.1.2] — 2026-07-12

Fixed

• inbox/ is now a proper reserved tier in walk()/load() (tools/lib/okf.mjs), on the same
  footing as secret//mirror/: excluded by default, opt-in via includeInbox /
  --include-inbox (okf-query, okf-recall). Before this fix, the first ever write through
  memory_write_inbox (MCP) — whose frontmatter carries only okf_version, no type — made
  samemind query validate permanently non-conformant, because raw inbox notes were treated as
  ordinary graph concepts. (#4)
  • validate/list/links/rel/get no longer see inbox/ by default.
  • tools/consolidate.mjs keeps reading inbox/ (opts in explicitly — that's its whole
    purpose: mapping raw notes to canon gaps).
  • MCP memory_search/memory_get/memory_list never returned inbox content and still don't.
  • Added a regression test: a fresh bundle → memory_write_inbox → validate stays conformant.

[0.1.1]

Fixed

• npx/.bin symlink: resolve argv[1] through realpath so the CLI's isMain check
  recognizes itself when invoked via a symlinked bin (e.g. npx samemind).

[0.1.0]

• Initial release.

v0.2.0 — The Flywheel

Changelog

All notable changes to this project are documented in this file.
Format loosely follows Keep a Changelog.

[0.2.0] — 2026-07-12 «The Flywheel»

Added

• Knowledge-cycle layer — three new concept types close the loop from facts to plans
  (docs/knowledge-cycle.md): Analysis (facts → pattern → implications), Research
  (question → findings → verdict, with source citations), Idea with a maturity status
  (spark → incubating → adopted / rejected, rejected_reason required). Edge conventions
  over existing relations: informs (Analysis/Research → Idea), spawned_by
  (Research → Analysis), led_to (Idea → Plan).
• Ideas on the board — samemind board gains a 💡 Ideas section: incubating first,
  then sparks; adopted collapse into an "Adopted → Plans" line via led_to; rejected hidden.
• Agent reflection protocol — memory-protocol and all three snippets teach agents to
  write reflections on immature Ideas through their own inbox (target: <idea path>),
  never editing the idea file directly; curation merges into ## Reflections.
• HTML projections — samemind board --html [--out <file>] and
  samemind handoff --html: self-contained pages (inline CSS, light/dark via
  prefers-color-scheme, zero JS, zero external resources) with code-generated SVG
  visualizations (kanban bars, ideas strip, decisions timeline). Markdown stays the canon;
  HTML is always a generated projection. Board/handoff internals split into
  model + renderers so both outputs share one data path.
• Validator warnings for the new types (Idea without status, rejected without
  rejected_reason); scaffold templates for all three types in samemind init;
  demo bundle gains a linked Analysis → Research → Idea working example.

[0.1.2] — 2026-07-12

Fixed

• inbox/ is now a proper reserved tier in walk()/load() (tools/lib/okf.mjs), on the same
  footing as secret//mirror/: excluded by default, opt-in via includeInbox /
  --include-inbox (okf-query, okf-recall). Before this fix, the first ever write through
  memory_write_inbox (MCP) — whose frontmatter carries only okf_version, no type — made
  samemind query validate permanently non-conformant, because raw inbox notes were treated as
  ordinary graph concepts. (#4)
  • validate/list/links/rel/get no longer see inbox/ by default.
  • tools/consolidate.mjs keeps reading inbox/ (opts in explicitly — that's its whole
    purpose: mapping raw notes to canon gaps).
  • MCP memory_search/memory_get/memory_list never returned inbox content and still don't.
  • Added a regression test: a fresh bundle → memory_write_inbox → validate stays conformant.

[0.1.1]

Fixed

• npx/.bin symlink: resolve argv[1] through realpath so the CLI's isMain check
  recognizes itself when invoked via a symlinked bin (e.g. npx samemind).

[0.1.0]

• Initial release.

v0.1.2 — inbox is a proper tier

Changelog

All notable changes to this project are documented in this file.
Format loosely follows Keep a Changelog.

[0.1.2] — 2026-07-12

Fixed

• inbox/ is now a proper reserved tier in walk()/load() (tools/lib/okf.mjs), on the same
  footing as secret//mirror/: excluded by default, opt-in via includeInbox /
  --include-inbox (okf-query, okf-recall). Before this fix, the first ever write through
  memory_write_inbox (MCP) — whose frontmatter carries only okf_version, no type — made
  samemind query validate permanently non-conformant, because raw inbox notes were treated as
  ordinary graph concepts. (#4)
  • validate/list/links/rel/get no longer see inbox/ by default.
  • tools/consolidate.mjs keeps reading inbox/ (opts in explicitly — that's its whole
    purpose: mapping raw notes to canon gaps).
  • MCP memory_search/memory_get/memory_list never returned inbox content and still don't.
  • Added a regression test: a fresh bundle → memory_write_inbox → validate stays conformant.

[0.1.1]

Fixed

• npx/.bin symlink: resolve argv[1] through realpath so the CLI's isMain check
  recognizes itself when invoked via a symlinked bin (e.g. npx samemind).

[0.1.0]

• Initial release.