v0.9.1

[0.9.1] - 2026-07-02

Added

• MCP Bundle (.mcpb) release artifact — the release workflow now assembles a
  single palace-<version>.mcpb bundle containing the Linux x86_64, macOS
  arm64, and Windows x86_64 binaries plus an MCPB manifest.json (binary server
  type with per-platform overrides). The bundle and its SHA-256 are uploaded as
  release assets, enabling one-click install in MCPB-aware hosts.
• Official MCP registry auto-publish — a server.json (namespace
  io.github.ancientice/palace-rs, MCPB package pointing at the release bundle)
  is published to registry.modelcontextprotocol.io on every tagged release. The
  release workflow authenticates via GitHub Actions OIDC (no stored secret) and
  fills in the version, download URL, and bundle hash automatically.

v0.9.0

[0.9.0] - 2026-06-29

Added

• Ambient warm-start recall at session start — the SessionStart hook now
  injects real recalled memory alongside the protocol text: recent diary entries
  for the session's project (cross-agent, so a different agent the next day still
  benefits) plus the top drawers of the wing the cwd maps to. Prior
  investigations land in context deterministically, without the agent having to
  call any palace tool. Fail-open: a missing or empty palace yields the protocol
  text alone. Applies to every client with a SessionStart hook (Claude Code,
  Codex, Cursor).

Fixed

• Read recall gap — PostToolUse auto-recall fired for the Read tool (it is
  in the installed matcher) but never surfaced anything, because the query
  extractor did not read Claude's file_path field. Reads now surface relevant
  prior memory. The Cursor postToolUse matcher was also aligned to
  Grep|Read|Glob for parity with Claude.
• UTF-8 truncation panic in memory layers — snippet truncation in the L1/L2/L3
  layer renderers sliced on a byte index (&s[..197] / &s[..297]), panicking
  when a multibyte character (e.g. ─, —) straddled the boundary — reachable
  via palace wake-up, palace recall, and palace search on real mined
  content. Truncation is now char-boundary safe.

v0.8.0

[0.8.0] - 2026-06-19

Added

• First-class protocol tools — the nine memory-protocol-critical MCP tools
  (palace_status, palace_session_context, palace_diary_search,
  palace_project_status, palace_search, palace_kg_query,
  palace_preference_search, palace_diary_write, palace_kg_add) now emit
  _meta."anthropic/alwaysLoad" = true in tools/list. Clients that honor the
  hint (Claude Code >= 2.1.121) keep these resident at session start instead of
  deferring them behind tool search, so the mandatory three-trigger protocol
  (SESSION START / BEFORE ANSWERING / AFTER WORK) no longer depends on the agent
  remembering to load the tools first. All other tools remain deferrable to keep
  standing context cost low.

Fixed

• Docs — updated remaining contributor, mission, and Claude notes references
  from the old mempalace-rs / mempalace mcp names to palace-rs /
  palace mcp.

v0.7.0

[0.7.0] - 2026-06-13

Added

• First-class wings registry — a new wings table is now the source of truth
  for every wing (both code repos and non-repo topics), recording kind
  (project/topic), description, project_path, and last_mined_at. It is
  populated idempotently on every database open by backfilling distinct wings
  already present on drawers (wing_diary__*/general/conversations are
  inferred as topics, everything else as projects), so existing palaces upgrade
  automatically without touching drawer data.
• MCP project-awareness and on-demand mining — three new MCP tools:
  palace_project_status (reports whether the current repo is mined /
  registered_not_mined / unknown, with a recommendation), palace_mine
  (mines a code repository on demand after the user agrees, auto-initialising
  palace.yaml for first-time repos and running with its own short-lived
  connection so the server's shared connection is untouched), and
  palace_create_wing (declares a topic wing for non-repo users such as sales
  or PMs). palace_remember and palace_add_drawer now auto-register an unknown
  wing as a topic, and palace_list_wings returns full registry records.
• palace wings CLI subcommand — lists registered wings with their kind,
  drawer counts, and last mined time.
• Session-start PROJECT CHECK guidance — the bundled memory protocol now
  tells agents to call palace_project_status when entering a workspace, ask the
  user before mining a repo (mining is repo-only and implicitly initialises it),
  and offer palace_create_wing for non-repo topics.

Changed

• Mining records registry state — miner::mine marks the wing as mined
  (set_wing_mined) on success, the watcher refreshes mined state, and mine
  gained a quiet flag so MCP-triggered mining does not write progress to the
  JSON-RPC stdout stream.

v0.6.1

[0.6.1] - 2026-06-13

Removed

• Dead llm module and --no-llm flag — the unused local LLM refinement
  scaffolding (src/llm.rs), its pub mod llm; export, the empty llm Cargo
  feature, and the palace init --no-llm flag have been removed. Nothing in the
  default path used them.

Fixed

• No more panic on an invalid built-in entity pattern — compile_regex in
  the entity detector no longer panic!s if a built-in regex fails to compile;
  it logs to stderr and falls back to a never-matching pattern so detection
  degrades gracefully instead of crashing. Covered by new tests that verify the
  built-in patterns compile and the fallback behaves.

Changed

• Crate metadata — removed the broken docs.rs/palace-rs documentation link
  (the crate is not published there) and pointed homepage at
  https://palacememory.com.
• Docs — corrected stale MEMPALACE_* / mempalace.yaml migration claims in
  the README (those names were removed in 0.3.0) and linked palacememory.com
  for remote mode.
• CI/release — the install smoke matrix now targets aarch64-apple-darwin
  to match the actual macOS runner architecture, and the release workflow runs an
  install-smoke job that downloads the freshly published asset on macOS and
  Linux and verifies an end-to-end install.

v0.6.0

[0.6.0] - 2026-06-09

Added

• Remote MCP mode (shared palace-server client) — palace mcp can now run as
  a transparent stdio→HTTP bridge to a shared remote palace-server instead of the
  local palace. Every AI client keeps its existing stdio registration; when remote
  mode is on, each JSON-RPC request is forwarded to the server's /mcp endpoint
  with a Bearer API key and the response is streamed back verbatim
  (application/json and text/event-stream both handled). This lets a whole team
  share one memory backend in their own infrastructure without per-client URL/header
  wiring.
  • palace remote set --endpoint <url> [--api-key <key>] — store the remote
    endpoint and ps_… API key. The endpoint accepts a bare host, a base URL, or a
    full /mcp URL (normalised automatically). Omit --api-key to be prompted on
    stdin so the key never lands in shell history.
  • palace remote on / palace remote off — switch the MCP server between the
    remote palace-server and the local palace.
  • palace local — alias for palace remote off.
  • palace remote status — show the current MCP mode, normalised endpoint, and
    masked API key.
  • palace remote test — one-shot connectivity + auth probe that runs the MCP
    initialize handshake and reports the number of tools the remote exposes.
  • New configuration surface: PALACE_MCP_MODE, PALACE_REMOTE_ENDPOINT, and
    PALACE_API_KEY environment variables, plus the mcp_mode, remote_endpoint,
    and remote_api_key keys in ~/.palace/config.json (written with owner-only
    0600 permissions since the file holds the API key).
  • For non-standard responses, the proxy unwraps palace-server's text envelope for
    initialize and tools/list so any spec-compliant MCP client understands the
    handshake, while passing tools/call and error responses through unchanged.

v0.5.0

[0.5.0] - 2026-06-02

Added

• Automatic memory hooks (all clients) — palace install now registers
  user-scope session start, post tool use, and stop hooks for every client
  that supports them, enforcing memory use in every project without per-project
  rules. post tool use auto-recalls relevant memory while the agent
  investigates; stop nudges the agent to record its work when it engaged Palace
  but saved nothing. Coverage:

  • Cursor — ~/.cursor/hooks.json (recall matched to Grep/Read).
  • Claude Code — ~/.claude/settings.json nested hooks (recall matched to
    Grep/Read/Glob).
  • Codex — ~/.codex/hooks.json nested hooks (recall matched to Bash,
    since Codex investigates through the shell; run /hooks once to trust them).
  • Claude Desktop — no hook system, so it remains rules-only (CLAUDE.md).

  Claude Code and Codex share a "Claude-style" output dialect
  (hookSpecificOutput.additionalContext; decision: "block" + reason on
  stop), selected via palace hook <event> --client <cursor|claude|codex>. The
  palace::hooks module exposes the pure, client-aware response builders.

• Cross-agent recall — palace_diary_search accepts all_agents and
  project_path to surface investigations recorded by any agent, and
  palace_session_context falls back to another agent's recent project work when
  the caller has none, so a different agent the next day still benefits from
  prior work.

Changed

• Write-path dedup — palace_diary_write (scoped per agent) and
  palace_remember now skip near-duplicate content, preventing diary/fact bloat
  from repeated similar writes. Durable decisions still belong in the
  deduplicated knowledge graph.

Fixed

• Resilient model download — the embedding model fetch now retries with
  exponential backoff, so transient HuggingFace failures (notably HTTP 429 rate
  limiting) no longer fail the first embedding call or CI. CI also caches the
  downloaded model (~/.cache/huggingface) across runs to avoid repeated
  downloads.

v0.4.0

[0.4.0] - 2026-05-22

Added

• Preference lane v2 — preference-shaped drawers now store a
  preference_span metadata field and optional pref_embedding secondary
  vector. Preference queries receive a separate preference_match score so the
  known single-session-preference weakness can improve without adding an LLM
  to the default path.
• Query intent provenance — searches classify queries as preference,
  decision, how_to, definition, temporal, or unknown, and return the
  intent in MCP and Palace::search_with_provenance results.
• Optional local rerank path — PALACE_RERANK=1, palace search --rerank,
  or palace_search({"rerank": true}) reranks top hybrid candidates and returns
  rerank_score provenance. The default search path is unchanged.
• Gain v2 feedback — palace_gain now accepts an optional record payload
  for explicit usefulness feedback and reports precision_at_1,
  precision_at_5, and per-intent precision. Diary entries that cite returned
  drawer IDs also infer useful feedback automatically.
• Eval CI lane — CI now runs the coding-agent memory eval separately and
  checks the sampled LongMemEval baseline guard.

Changed

• palace upgrade-embeddings --refresh-preferences and
  palace_upgrade_embeddings({"refresh_preferences": true}) can refresh
  preference-span embeddings while re-embedding drawers.
• Palace::search_with_provenance is additive: results include
  preference_match, intent, and rerank_score.

v0.3.2

[0.3.2] - 2026-05-15

Changed

• Agent rule restructured into 3 hard trigger blocks — RULE_BODY (installed
  into .cursor/rules/palace.mdc, .codex/AGENTS.md, .claude/CLAUDE.md) and
  PALACE_PROTOCOL (embedded in palace_status MCP response) are now structured
  as three imperative blocks: SESSION START, BEFORE ANSWERING, and
  AFTER WORK. The previous 11-step and 9-step numbered lists caused agents to
  front-load steps 1–2 and deprioritise the rest; the new block layout makes each
  trigger an unconditional gate, improving palace_diary_search and
  palace_kg_query call rates in practice.
• rule_is_weak detection updated to check for the new block keywords
  (SESSION START, BEFORE ANSWERING, AFTER WORK) rather than the old numbered
  step phrases. Run palace install to refresh installed rules.

v0.3.1

[0.3.1] - 2026-05-15

Added

• palace_import MCP tool — import drawers from a JSON snapshot produced
  by palace_export. Skips drawers that already exist (idempotent). Returns
  inserted, skipped, and total counts.
• palace_upgrade_embeddings MCP tool — re-embeds all drawers using the
  current embedding model. Run after upgrading the model to keep search quality
  consistent.
• palace_prune MCP tool — deletes drawers filed more than older_than_days
  days ago. Irreversible; export first if unsure.
• palace prune --older-than-days N [--dry-run] CLI command.
• palace upgrade-embeddings CLI command.
• palace timeline [--entity <name>] CLI command — human-readable
  chronological view of knowledge-graph facts, optionally filtered to one entity.