Skip to content

v0.23.0 — Recall you can trust, plus selective project handoff

Choose a tag to compare

@Paul-Kyle Paul-Kyle released this 14 Oct 02:59
· 11 commits to main since this release

[0.23.0] — 2026-10-14

Compatibility: a same-slug re-save with update_policy: replace, or without a supplied or inherited append policy, now keeps the record's existing frontmatter (custom keys, title, status, entities, links), replaces the body, and clears metadata tied to the old body (description, summary, claims, sources); a caller that relied on the old reset must send the keys it wants changed, or an explicit empty title, contradicts, backed_by or external_refs to clear one. Rewrites keep the existing visibility, scope and access unless the save supplies them (supplied values are validated as before), and a rewrite over frontmatter that does not parse is refused. palinode init --force now replaces an existing Palinode block in an instruction file instead of appending a second one. A search date_after / date_before that is not ISO-8601 is now a 422 instead of being ignored. The OpenClaw plugin openclaw-palinode 0.23.0 requires OpenClaw 2026.10.3 or later.

Recall you can trust, plus selective project handoff. A project's reviewed decisions and insights can move between stores (palinode handoff export / import) with a preview on both sides, and an import can never overwrite, edit or elevate anything. Automatic recall is more precise and honest about weak matches, the OpenClaw plugin's automatic recall reaches the agent again, rewrites no longer drop a record's privacy fields, and this release carries fixes from seven outside contributions.

OpenClaw verification: no live OpenClaw gateway check was run for this release. Compatibility was checked against the real OpenClaw 2026.10.3 SDK declarations, and all 165 plugin tests passed.

Added

  • List filters on frontmatter with where. Each entry is key=value and every entry must match: a scalar by its text, a boolean as true/false; a list when any element matches; a record without the key, or with a null or a mapping there, never matches. where treats only dashed ISO dates (2026-10-12, optionally with a time of at least hours and minutes) as dates, and reads them as the search date window does: a timestamp matches the same instant whether written with T or a space, with Z, an offset or none, a value without an offset is UTC, a stored bare date is midnight UTC, and a bare date in where matches any time on that UTC day. 20261012, 2026-W42 or 2026-10-12T10 are compared as text, because a front-matter value such as an id can look like one. On GET /list it is a repeated query parameter, on the CLI a repeatable palinode list --where, and on the MCP palinode_list tool an array. A malformed entry is a 422 on REST, and exit status 2 on the CLI or an Error: reply on MCP before any request is sent (#238, public PR #244, thanks @marcin-rogalski).

  • A save can return before its embed, and the save request timeout is configurable. On a slow embedder, POST /save could outlast the CLI and MCP clients' fixed 30 s timeout after the write had landed, so the client reported a failure for a save that succeeded and a retry saved it twice. embeddings.defer_embed_on_save: true (env PALINODE_DEFER_EMBED_ON_SAVE, default off) makes the save use the existing cold-embedder path: keyword-searchable rows now, embedded: false with the deferral in index_error, and vectors on the watcher's next pass. PALINODE_SAVE_TIMEOUT (seconds, default 30) sets the CLI and MCP save request timeout; a malformed or non-positive value keeps 30 s with a warning. Both default to the previous behaviour (#234, public PR #235, thanks @marcin-rogalski).

  • palinode handoff export: a previewed, selective export of project memory. Writes one JSON bundle of a project's current Decision and Insight records, or the records you name, for a teammate or another agent's store. The preview lists every record that would leave, the frontmatter that won't (core, authority, update_policy, scope, visibility and every other field outside a fixed allow-list), quotes of other records, text addressed to AI agents, the records left out and why, and a digest. The bundle is written only with --expect <digest>, and only if nothing changed since the preview. Private and restricted records are never exported. people/, daily/, inbox/, core, retired and other-project records are left out unless you name them. Lineage, evidence and conflict links are followed one hop, and retired dependencies travel retired. The bundle counts left-out records but doesn't name them. CLI and POST /handoff/export only; there is no MCP tool. See docs/HANDOFF.md.

  • palinode handoff import: a previewed import of a handoff bundle that can't overwrite, edit or elevate anything. The preview shows where each record would land, which are already in the store (the same origin record and content, from any earlier bundle), which are rejected and why, the frontmatter the bundle carried that is ignored, text addressed to AI agents, and a digest. --apply --expect <digest> writes the records only if neither the bundle nor the store changed since the preview, as new files in one commit naming the bundle. A taken path lands renamed (<slug>-handoff-<id>.md) with links from the rest of the bundle remapped; existing records are never edited. Only the fields an export sends are read, so a hand-edited bundle can't add core, authority, scope, visibility or source. Every imported record is marked source: handoff-import with an imported_from block. Records are rejected for an unsafe path (including a directory that resolves outside the store), a body that doesn't match its hash, size limits, the save path's security scan, or tool-call markup. --project maps the sender's project to yours; without it, an unknown project is refused. An interrupted apply is committed by the next one. A capture pause stops the apply, not the preview. CLI and POST /handoff/import, POST /handoff/import/apply only; there is no MCP tool.

  • palinode hooks update and the hooks_current doctor check. palinode init copied the Claude Code hook scripts once and never refreshed them, so hook fixes in a release never reached an existing install. Each packaged hook now carries a # palinode-hook-version: X.Y.Z stamp, and init reports the version and sha256 of each hook it installs. palinode doctor walks the Claude Code settings cascade (user, project, project-local) and warns when a registered palinode hook differs from the packaged copy, naming the refresh command. palinode hooks update replaces each stale copy, keeps the old one as <name>.bak-<timestamp>, installs registered-but-missing hooks, never touches other hooks or symlinked ones, and writes nothing on a re-run; --dry-run shows the diff. Upgrading: run palinode hooks update (from each project that registers hooks in its own .claude/settings.json) after installing this release.

Changed

  • Same-slug body replacements keep record frontmatter. update_policy: replace and saves with no policy retain omitted record keys, including custom fields, entities, expiry, and links. Old-body descriptions, summaries, claims, and source anchors are cleared so enrichment can describe the new body, and a stale_backing flag still clears, since a re-save is the re-verification that consolidation asks for. Caller-supplied values still win.

  • openclaw-palinode 0.23.0 supports OpenClaw 2026.10.3. Moved to the SDK core subpath, aligned tool results and prompt injection with the typed hooks, and declared startup and CLI ownership in the plugin manifest.

  • Auto-description and auto-summary prompts state a word budget instead of a character count. In a measurement with a local Qwen3-4B, 68% of generated descriptions overshot the 130-character rule and were clipped with "…". The prompts now ask for at most 16 words (descriptions) and auto_summary.max_chars // 8 words over one or two sentences (summaries), and tell the model to lead with the subject rather than the note type ("A project snapshot of …"). On a 12-note sample, descriptions that fit the cap went from 3/12 to 10/12. No config change is required; auto_summary.max_chars still sets the summary cap.

  • Automatic isolation-only recall notices are delivered once per session. The Claude Code hook and shared Pi/Cline core suppress repeated confident bundles whose only news is other-project withholding. Content-bearing bundles retain the line, and explicit resolve/search keep it on every request.

  • README: current-release wording for v0.22.0. Drops the pre-release "v0.21-capable source checkout" phrasing for lexical mode and client fragments, and says how a remote HTTP client keeps recall scoped (palinode mcp-config --http --project <slug>), since a request without the X-Palinode-Project header is unscoped.

Fixed

  • A record the caller may not see no longer sets a search floor. Each arm's relative floor (search.vector_relative_floor on the vector arm, search.fts_threshold or search.lexical_fts_threshold on the keyword arm) was measured against the best candidate the arm fetched, and the visibility gate ran afterwards. A private or restricted record, or under project isolation a record from another project, could set the floor, push out a visible match, and then be withheld itself, so the search came back with fewer results or none. POST /search now measures each floor against the best candidate the caller may see, using the same visibility check as the gate, in hybrid, vector-only and lexical modes. Records withheld for another project are still counted in other_projects_withheld (#245). New tests/test_relative_floor_over_visible_hits.py (real SQLite + tmp_path, through POST /search) (#245, public PR #253, thanks @marcin-rogalski).

  • The search date window handles the calendar's edge and every date-only bound. A last_updated stamp whose UTC reading falls outside years 1 to 9999 (9999-12-31T23:59-01:00, 0001-01-01T00:00+01:00) raised OverflowError out of ranker.parse_iso_utc, so one such record made every search with date_after or date_before fail, and the same value as a bound answered 500. It now reads as not a timestamp: the record is kept by the window, like any stamp that cannot be read, and the value as a bound is a 422. A date_before in a date-only form other than YYYY-MM-DD (20261012, 2026-W42-1) was midnight, so it dropped the rest of its day; every date-only form the ISO parser accepts now covers its whole UTC day. where on GET /list is unchanged (#243, public PR #252, thanks @marcin-rogalski).

  • Automatic recall through the OpenClaw plugin reaches the agent again. Since the change that moved recall to system-role context, the plugin's before_prompt_build hook returned a systemContext field, which OpenClaw has never accepted (its prompt-mutation fields are systemPrompt, prependContext, appendContext, prependSystemContext and appendSystemContext). OpenClaw dropped it silently, so autoRecall injected nothing. The hook now returns prependSystemContext, keeping the recalled memory in the system role. Because that changes the system prompt every turn, a provider's prompt cache no longer covers it.

  • palinode init --force replaces instruction-file memory blocks in place. On .claude/CLAUDE.md, AGENTS.md, and .cursor/rules/palinode.md, a forced re-run now replaces a complete Palinode block while preserving bytes outside it. Complete duplicate blocks collapse to one; an incomplete block is refused with a manual repair instruction. Dry runs report “would replace” for existing blocks.

  • Doctor no longer calls a weekly that keeps losing the lock healthy. On the dogfood host the weekly and nightly cron entries fired in the same minute, the nightly won the consolidation lock every Sunday, and consolidation_last_run still reported "0 consecutive failures" with the last weekly success 290 h old. Cron now records a lock-refused attempt (lock_refused) and doctor reports it as a missed run for that cadence, separately from failed passes. Doctor also warns when a cadence's last success is stale: older than 1.5 cadences, or, with the activity gate on, older than the gate's ceiling (auto_gate.max_hours_elapsed) plus one cadence, since a quiet store may legitimately skip passes up to that ceiling. consolidation_schedule_effective flags a daily nightly and a Sunday weekly scheduled at the same minute. The cron.py example now schedules the weekly at 12:00.

  • OpenClaw migration preserves existing memories on a path collision. A section whose heading produced an occupied memory path could replace that file's body and front matter. The importer now writes different content to a deterministic renamed path with its own id, checks that path for collisions too, and reports renames in the result, CLI summary and migration log. Duplicates are now detected by body, not whole file: the frontmatter's per-run last_updated stamp had made every re-run look new, so a repeated migration now skips sections it already imported instead of copying them.

  • ActionItems can be listed. They are saved under inbox/, which GET /list always skipped. include_inbox=true (palinode list --include-inbox, include_inbox on the MCP palinode_list tool) lists it too; the default is unchanged (#238, public PR #244, thanks @marcin-rogalski).

  • Git CLI commands emit JSON when piped. palinode blame, including --claims, palinode rollback, and palinode push now use the shared TTY-aware formatter, so non-interactive output parses as JSON and keeps bracketed source text intact, while the interactive claims view still renders as prose. Interactive rollback and push show formatted output instead of a raw Python dict (#210, public PR #242, thanks @1cbyc).

  • Layer splitting preserves an existing status file. Its prior frontmatter, including custom keys and status, survives; its previous status body moves into history before the new current status is written.

  • An append save keeps the record's front matter. A save that resolved to appended rebuilt the front matter from that save alone, carrying forward only created_at, update_policy and epistemic. A record saved with a title, a status, core and entities, then appended to by a save that sent only the new content, kept both blocks of body and lost every one of those keys. Now every prior key the append does not set carries forward, custom keys included, and a key the append sets wins, an explicit core: false included. Keys every save derives (id, category, type, content_hash, last_updated, auditable_memory, source) are never carried. expires_at carries unchanged, so an append cannot move expiry, unless the append sends a ttl or an expires_at of its own. entities from the append win when non-empty; an empty or missing list keeps the prior one, and a project the append names joins it. The ## See also footer lists the string items of the entities the front matter ends with. A carried description or summary stays, so no enrichment is re-queued; send description: "" to have it regenerated. With ku_compat on, ku_version is re-stamped, and lifecycle is re-derived when the append sets status. Other replace frontmatter and first saves are unchanged (#236, public PR #237, thanks @marcin-rogalski).

  • Search date filters compare instants, not strings. date_after / date_before (and the since_days vs date_after pick) were compared to last_updated as text, so a bound with a UTC offset (2026-10-12T20:00:00-05:00), a space separator, or a bare-date date_before landed on the wrong side of a record. Both sides are now parsed to UTC by one helper, ranker.parse_iso_utc (a value without an offset is UTC, a bare-date date_before covers that whole day), on the vector, hybrid/lexical and recency-list paths. A bound that is not ISO-8601 is now a 422 instead of a silent string comparison (public PR #240, thanks @vishalhabib99).

  • CHAT and consolidation fallbacks no longer go down with the primary. The circuit breaker was keyed by role alone, so the primary and every llm_fallbacks entry shared one breaker: once a dead primary opened it, each fallback was fast-failed too and the whole chain was skipped exactly when it was needed (and a dead fallback could open the primary's breaker). A call whose base_url points at a host other than the role's configured one now gets its own breaker. The role breaker, which palinode doctor reports, still tracks the configured host.

  • The CHAT fallback budget no longer counts free local hosts. auto_summary.llm_fallback_max_per_run bounds Sonnet egress, but it was spent on every escalation, including ones a local model answered, so a primary outage capped enrichment at 10 files per backfill run. Fallback entries now take metered (default true); an entry with metered: false neither spends nor is blocked by the budget, and an exhausted budget skips only metered entries.

  • CHAT fallbacks use the caller's timeout instead of 60 s. Descriptions now give each fallback describe_timeout_seconds and summaries 30 s, so a hung backup fails within the breaker's 30-second window and its own breaker can open.

  • A keyword match that chance would produce no longer makes a delivery confident. Short commands such as "write the release notes for this" reached confident on the keyword arm alone: one repeated incidental term, or a few everyday terms a large store of work notes uses everywhere, clears the keyword mark with nothing about the request in the row. Each keyword row now carries how many chunks chance alone would give its combination of query terms, and the keyword arm claims confident only on a row below one; otherwise the delivery is weak with co_occurrence: "chance" in its retrieval block. One-term queries (exact identifiers), rows whose every matched term is in half the index or more (a store of a few records on one subject), and the keyword arm's corroboration of the vector arm are unchanged. On the relevance benchmark's prompt set, extended to 34 answerable and 40 no-answer prompts with short operations and coding requests, hybrid automatic delivery goes from 31/34 answerable and 2/40 no-answer to 31/34 and 1/40; the published question set's verdicts are unchanged.

  • Importing the API server, the watcher or the MCP server no longer creates directories in the store. The operations log handler opened <memory_dir>/logs/operations.jsonl at import, and the MCP audit logger and the API's retrieval logger created .audit/ when they were built, so anything that merely imported these modules (a CLI subcommand, a test collecting a module) wrote into whatever store PALINODE_DIR or ~/palinode named. The directories and the log file are now created on the first write, and the API's CORS and bearer-auth status lines are logged at startup instead of at import.

  • The CLI honours PALINODE_API_URL, the variable the hooks and plugins read. A client set up for the hooks used to get a CLI that ignored it and fell through to the server's bind address, usually loopback, failing with "connection refused". PALINODE_API still takes precedence.

  • Session-start "Recent decisions" have a stable order when files share a timestamp. Decisions written within one timestamp tick have the same mtime, and their order then came from directory scan order, which varies by filesystem and run. The file path now breaks ties.

  • A confident resolve that the budget emptied no longer says nothing answers. When every matching record is refused for budget, the bundle says "Matches did not fit the budget." instead of "Nothing in memory answers this." on every surface. The per-turn hook and shared plugin core stay silent on such a bundle rather than delivering it as the once-per-session other-projects notice; the notice is still delivered when nothing in scope answers. A completed resolve still never falls back to search.

  • Session-end hook topics after /clear are the human's first ask, not harness markup. After /clear a transcript opens with a <local-command-caveat> turn, so automatic captures read "Topic: Caveat: The messages below…". The hook (examples/hooks/palinode-session-end.sh and the identical palinode init copy) now drops the caveat block whole, skips harness-generated (isMeta) turns and bare argument-less slash-command invocations, reads only text blocks from block-array content, and takes the first user turn that still has human text.

  • palinode doctor reports unreplayed session-end fallback records. A new session_floor_fallback check warns when the current project holds a non-empty .claude/session-floor-fallback.jsonl, with the count, when it was last written, and the palinode session-end replay step. These records were previously never surfaced.

  • Resolve keeps current assertions within hook budgets by bounding query-relevant unlinked evidence and delivering a qualified minimal form when needed; full discovery remains available in structured responses and explicit search.

  • Resolve selection accounts for retired seed windows. Bounded re-fetching lets current records below retired query seeds reach resolution, while preserving successor traversal. Fully empty query bundles explain lifecycle or budget withholding in their text and retrieval diagnostics; seed confidence is preserved.

  • Remote automatic recall gets a usable resolve deadline and withholds retired fallback hits. The per-turn hook and shared plugin core default to 1000 ms for remote URLs and retain 250 ms for loopback, with the existing environment override. Automatic/passive search filters live lifecycle state before trimming and assessing delivery; explicit search keeps its existing behavior. The index-readiness diagnostic that every /resolve and receipt search computes no longer probes the vector table once per chunk: it took about 2.7 s on a 12,000-chunk store, above any per-turn deadline, and now takes tens of milliseconds.

  • Per-prompt recall precision. Resolve bundles and receipts expose confidence and pre-fusion arm scores on API, MCP and CLI. Automatic recall in the Claude Code hook and shared plugin core shapes raw prompts on the server (harness markup and small talk are dropped; questions and coding requests are kept) and delivers only confident resolve or search results; weak, absent and missing verdicts produce silence. Trigger checks use the client’s project scope before consuming cooldowns. A synthetic prompt-shaped extension to the relevance benchmark measures answerable recall and no-answer delivery.

Removed

Security

  • Rewrites retain privacy frontmatter. Re-saving an existing slug with replace, no policy, or append carries its visibility, scope, and access unless explicitly supplied; new status and history layers inherit those fields from their source.

Full changelog for v0.23.0