doobidoo/mcp-memory-service v12.7.0

on GitHub

Special Thanks

Thanks to the external contributors in this release: @ShahadatTest (Credentials tab requests, #1841), @massimiliano1991 (typed-edge gates, #1888), @mbreitkreuz (conflict tags and entity linking, #1872, #1889), @tonydzi (typed edges off by default, #1887), @qwist1233-cpu (untagged deletes and dashboard translations, #1871, #1822) and @ducanhnguyen223 (Codex short replies, #1716).

Highlights

• Behavior change: MCP_TYPED_EDGES_ENABLED now defaults to false. Consolidation stores inferred associations as related instead of fixes / contradicts. Set it to true to keep the old behavior; the stricter #1457 gates apply then.
• Store scoping is enforced in the storage layer for metadata updates, the fallback=true search and beliefs.
• conflict:unresolved no longer outlives the conflict, and entity linking no longer overwrites contradicts edges.
• New opt-in contextualCapture setting for Codex short replies.

Changelog

Fixed

• Consolidation's typed-edge gates are the ones its docs state (#1457). The docs said
  confidence 0.75 and similarity 0.65, while the code shipped 0.50 and 0.45 and consolidation
  ran at 0.6 and 0.45. On three hand-checked stores neither setting made a label right
  (0 of 139), but the #541 gates write about ten times fewer of them, so those are now the
  constructor defaults and the consolidator passes them explicitly. The 0.65 is on the
  normalized score (cos + 1) / 2, i.e. cosine 0.3, and the docstrings say so. These gates
  apply only with MCP_TYPED_EDGES_ENABLED=true, off by default since #1887.
  scripts/maintenance/update_graph_relationship_types.py builds the same engine, so it
  picks up the 0.75; it passes no similarity, so the similarity gate does not apply there.
• Optionally capture short Codex replies with their preceding assistant context (#1664). The new contextualCapture setting is off by default and stores an opted-in reply/response pair with source turn metadata, without an LLM call.
• Search filter tooltips and most of the Documents tab stayed English (#1817).
  The five filter tooltips and the three Documents help icons already had catalog
  keys, but only on data-i18n-aria, so the tooltip a user actually hovers kept its
  English title; data-i18n-title now sits next to each aria attribute. Those icons
  also sat inside the element whose text is translated, and applyTranslations()
  writes textContent/innerHTML, so the first render dropped them from the page --
  the translatable text now lives in its own <span> and the icon stays beside it.
  The upload history (No uploads yet, its error line), the View/Remove buttons of
  both document lists, the drop zone that app.js rebuilds after an upload, the
  chunk-overlap illustration, the processing-mode help heading, the memory-viewer chunk
  count and the removal confirmation go through this.t() or a data-i18n* attribute
  now, with twenty-two new strings in all eleven catalogs. The chunk-size help icon also
  gets its own info-icon-chunk class: it shared the bare .info-icon hook that
  setupDocumentsEventListeners() queries, so once the processing-mode icon stayed in
  the page the wrong icon answered the click. documents.supported also lists DOCX and
  PPTX, which semtools_loader.py has accepted all along, and the static #fileInput
  now accepts .docx,.pptx like the input app.js rebuilds.
• Test Connection and Save in the Settings Credentials tab send their requests again (#1839). Both called apiCall() with a fetch-style options object instead of the method and the data as separate arguments, so neither request carried the method or body it needed. tests/web/test_credentials_apicall_args.py checks both calls.
• conflict:unresolved no longer outlives the conflict (#1840, mbreitkreuz). Superseding or deleting one side of a conflict (mark_superseded_batch, resolve_conflict, delete, a versioned update) left the tag on the other side, and resolve_conflict stripped it from the winner even while the winner still contradicted another memory. resolve_conflict also refuses a memory resolved against itself now. The partner's tag is now cleared in the same transaction once it has no current conflict partner left, a new version no longer inherits the tag, and the maintain conflicts step reports (dry run, the default) or clears tags left behind by older data, bulk deletes and sync. sqlite-vec only.
• POST /api/manage/delete-untagged now deletes through the storage layer on the SQLite backends instead of tombstoning the rows with a raw UPDATE on the shared connection. The new DeleteMixin.delete_untagged_memories() removes the graph edges, retracts the beliefs and summaries that cited the deleted hashes, appends one delete event per hash for delta-sync peers, and commits once under the connection lock — the same shape as delete_by_tags(). count_untagged_memories() joins it, so both endpoints now go through one predicate and the raw-SQL fallbacks are gone.
• The confirm_count the caller is held to is now re-checked where the rows are selected for deletion — inside the connection lock and the transaction on SQLite, before the first write — so a concurrent request that strips the last tag off another memory between the count and the delete can no longer widen the delete past what was confirmed. The endpoint answers such a race with the same 400 as a stale count.
• Korean landing page headline reworded (#1873). The hero read
  AI 어시스턴트에게 오래 남는 기억을; it now reads AI 어시스턴트를 위한 기억력,
  matching the page title.
• Entity linking no longer replaces contradicts edges (#1874, mbreitkreuz). With MCP_ENTITY_LINKING_ENABLED=true, storing a memory that contradicts an existing one recorded the conflict and then overwrote both contradicts edges with shares_entity, because a pair holds one edge and the linker runs after conflict detection. memory_conflicts and the maintain auto-resolve lost the pair while both memories kept conflict:unresolved. The linker now skips any pair that already has an edge other than has_entity / shares_entity, the rule rebuild_entities already used. Pairs overwritten before this fix are not repaired: if you ran with entity linking enabled, run memory_quality maintain as a dry run first. An overwritten pair shows up as two reported hashes joined by a shares_entity edge; resolve it with memory_resolve before running maintain with dry_run=false, which clears the tag and with it the last trace of the conflict.
• Store-scoped metadata updates are confined in the storage layer (#1875, part of #1683). update_memory_metadata and update_memory_versioned looked a memory up by content_hash alone. A claim on one store was only enforced by a get_by_hash pre-check in the MCP handler, so the check and the write were two separate steps. Both methods now take store, put AND store = ? into the lookup and the UPDATE, and report a memory in another store as not found without changing it. A metadata update whose UPDATE matches no row, because the memory was deleted or moved between the read and the write, now returns not found instead of success, and a versioned update aborts if the old row leaves the store while the new version is written. The MCP handler and PUT /api/memories/{hash} pass the claimed store through. None and "all" stay unscoped, so callers without a store behave as before. sqlite-vec and hybrid only; the other backends still refuse store-scoped updates until #1878 and #1879 land.
• memory_search with fallback=true now runs its exact and tag fallbacks in the same store as the main search. Before, both fallbacks searched the default store whatever store the call named, so a search in one partition could return another partition's memories. A per-store credential (read:<store>) can now make a fallback=true search; it used to need the global grant.
• Beliefs and the NLI on-store check respect the store (#1877, part of #1683). Beliefs aren't stored per store, so memory_search with include_beliefs returned beliefs from every store, and the opt-in NLI check on store (MCP_NLI_ON_STORE) compared a new memory with all of them. A per-store grant therefore could not cover include_beliefs. A belief is now visible in a store when none of its live source or contradicting memories sits in another store. Deleted memories don't count, so retention never hides a belief. The beliefs section of a search uses the same store as its memory results ("default" unless one is given), and the on-store check, in both the synchronous and the background path, compares a new memory only with beliefs visible in the store it was written to. A single-store user, with everything in "default", sees every belief and keeps challenging them as before. include_beliefs no longer needs the global grant.
• Consolidation no longer labels inferred edges fixes / contradicts by default (#1883).
  MCP_TYPED_EDGES_ENABLED now defaults to false. A typed label came from the pair of
  memory types (learning -> bug gave fixes, two note-like types gave contradicts),
  not from what the two memories say, and hand-checked samples from three stores found
  0 of 139 typed edges correct. Inferred associations are still stored, as related.
  This release does not relabel stored edges in bulk, but store_association() writes
  with INSERT OR REPLACE, so a pair that a later consolidation finds again is
  stored as related and loses its typed label. MCP_CONSOLIDATION_AUTO_SUPERSEDE
  acts on contradicts edges, so with the new default it has nothing to act on. To keep the old behavior, set
  MCP_TYPED_EDGES_ENABLED=true; the #1457 gates still apply.
• Keep selected source memories intact during semantic compression so rules retain
  their exceptions and conditions, including short caveats and separate paragraphs.
  Omit excerpts that exceed the summary budget instead of storing a partial rule.

Internal

• pre_pr_check.sh now runs the changelog fragment check. Check 6.11 calls
  scripts/ci/check_changelog_entry.sh --staged origin/main, the same script CI's
  changelog.yml runs, but against the index, so a missing fragment or one without a
  - list item fails locally instead of in CI after the push.

Full changelog: https://github.com/doobidoo/mcp-memory-service/blob/main/CHANGELOG.md