0.3.3

Memora 0.3.3

Search accuracy and hardening release.

Search

• Full-text search now queries the FTS index correctly, improving keyword relevance; substring matching remains only as an explicit fallback.
• limit is honored on memory_hybrid_search and memory_semantic_search (top_k still accepted).
• Searches and lists with lineage filtering (follow=active/latest) fill the requested result count even when top-ranked candidates are superseded, scanning beyond the previous 5,000-row window with a loud error at the safety bound instead of silent truncation.

Absorb & lineage

• Absorb updates supersede the current version of a memory, resolving through the supersession chain to the leaf.
• New classifier measurement harness: labeled fixture pairs, per-class precision/recall and confusion matrix, dry-run safe, with a --min-macro-f1 gate for regression testing.
• All LLM calls are bounded by an explicit timeout (MEMORA_LLM_TIMEOUT, default 60s). Measurement mode fails loud; runtime absorb degrades gracefully.

Tag policy

• The Cloudflare graph app validates tag writes (memory edit and chat) against a versioned policy stored per database, failing closed when the policy is unavailable.
• Wildcards support slash namespaces (memora/*) alongside dot namespaces; tags are capped at 100 characters, counted identically (Unicode code points) in Python and TypeScript and guarded by a shared conformance fixture.

Graph UI

• The WebGL canvas tracks its container through drawer transitions via ResizeObserver, fixing a sizing race under load.

CI

• New clean-install workflow: builds the wheel, installs into an empty environment, and runs the suite — on push, tags, and a daily schedule.
• New graph-ui workflow: browser tests for the graph UI (drawers, top bar, render-idle power behavior, database selector), tag-policy write tests, and lineage logic tests against a seeded local D1.

Docs

• Install instructions lead with PyPI (pip install memora-mcp); absorb, supersession lineage, and digest documented in Features; MEMORA_TAG_FILE format corrected (JSON array).

v0.3.2

memora 0.3.2

Consolidated 0.3.x release. Notes for the earlier 0.3.x tags remain in the repo as
RELEASE_NOTES_0.3.0.md and RELEASE_NOTES_0.3.1.md.

Fresh installs work again

mcp 2.0.0 removed mcp.server.fastmcp, and our dependency was an unbounded mcp>=1.0.0, so every
fresh install resolved to the new major and the server died at import with
ModuleNotFoundError. Now constrained to mcp>=1.0.0,<2 (the 1.x line is maintained in parallel).

The failure was invisible from both sides: a dead stdio MCP server looks identical to one exposing no
tools, and existing environments had mcp pinned to a working 1.x, so every test suite passed.

Reported and fixed by @BillyBunn in
#44.

Breaking

Embeddings. memora could be configured — following its own installer and README — into a state
where no embedding was ever computed and nothing said so. install.sh generated
openai/text-embedding-3-small while the README recommended OpenRouter as the base URL, and
OpenRouter serves no embeddings endpoint. Every call 404'd, one warning went to a log nobody reads,
and stores silently filled with TF-IDF keyword bags. In the store where this was found, 756 memories
had been keyword vectors for months.

• Every store needs one embedding rebuild. The model fingerprint now records backend, model,
  endpoint host and representation, so the old bare "openai" stamp no longer matches.
• Dense backends no longer fall back to TF-IDF. A provider failure raises instead of persisting a
  wrong vector. Set tfidf explicitly if you want it.
• MEMORA_EMBEDDING_API_KEY / MEMORA_EMBEDDING_BASE_URL are an atomic pair — set both or
  neither. A partial pair is rejected rather than borrowing the missing half from OPENAI_*, which
  could previously send one provider's secret to another provider's host.

Issues are no longer inferred. memory_create_issue and memory_create_todo are now the only
ways a memory becomes typed; memory_absorb and plain creates stay untyped knowledge. A keyword
classifier had mislabelled 130 knowledge memories as open issues. Existing typed memories are
untouched.

Added

• memory_verify_integrity — read-only doctor reporting representation counts, coverage and
  offending ids, with concrete remediation.
• MEMORA_EMBEDDING_STRICT=1 (recommended) turns a broken endpoint into a named error instead of
  silent degradation.
• Integrity derived from SQL and invalidated by a DB-owned change epoch, so an external writer cannot
  leave a stale "healthy" verdict behind.

Fixed

• Batch embedding responses are validated for cardinality, index coverage and uniform dimensions, and
  reconstructed by index — a reordered response could previously attach a vector to the wrong
  memory.
• Coverage uses indexed anti-joins in both directions; counting by subtraction let one orphan cancel
  one missing embedding.
• absorb no longer leaves untracked partial rows.
• Concurrent ensure_schema no longer raises on duplicate columns.
• install.sh and the README no longer generate or document the broken configuration.

memora-graph

• Supersession is visible. Superseded memories render dimmed amber with a SUPERSEDED badge and
  directed lineage arrows; a toggle collapses to current-state only. Only supersedes edges count as
  lineage — references/contradicts/implements were previously drawn as supersession. Half-written
  crossrefs are detected rather than shown as current, and the page reports
  LINEAGE UNAVAILABLE instead of a confident zero when it cannot tell.
• The 3D view no longer burns the CPU. Its render loop repainted 60×/sec forever, whether or not
  anything changed — measured at ~227% CPU on an idle page. The loop now stops once the layout
  settles and wakes on interaction: ~227% → 0.4%. Kill switch:
  localStorage.setItem("memora-graph.noIdle","1").
• Resizable timeline and detail drawers with independently persisted widths; per-tab panel widths in
  the default view.
• The database selector lists every configured database in both views (it was hardcoded in one).
• The top bar no longer hides behind open drawers.

Known issues

• Epoch time-of-check window — a query can use data that changed mid-call; the next call detects
  it. Transient, self-healing.
• memories_embedding_repairs is unbounded with no foreign key to memory lifetime; explicit id
  reuse can mark an unrelated row as recurring.
• D1 ownership recovery after a lost response lacks bounded retry and a unique operation record.
• External writers must populate the new columns. A writer emitting a thresholded/sparse encoding
  of a dense vector is reported as an encoding fault by id, and auto-rebuild is skipped deliberately.
• The default index.html view lacks force-graph's database-switch protection — after a switch or
  live refresh an already-rendered list can briefly show another store's authority state. Memory ids
  overlap across databases, so this is worth knowing.

Upgrading

1. Upgrade, then run memory_verify_integrity before anything else.
2. If it reports a repairable mismatch, run an embedding rebuild.
3. Set MEMORA_EMBEDDING_STRICT=1.
4. Verify by looking at a stored vector, not at your config — it should have your model's
   dimension count with numeric keys covering 0..N-1. Configuration that looks correct is what hid
   the original problem.

v0.2.29

Changes since v0.2.28:

• Add memory_digest MCP tool
• Tighten memory_digest aggregation and filtering
• Make memory_update metadata updates merge safely by default

v0.2.28

Notable changes since v0.2.27

Fixes

• MCP schema sanitizer (memora/server.py) — strips Pydantic-generated nullable anyOf combinators from tool input schemas before serving tools/list. Anthropic's tool validator rejected these, which 400'd the entire MCP handshake and made Claude Code disconnect the server. Reviewed end-to-end by an inter-agent bus loop (claude-B audit + codex review). 19 of 41 memora tools were affected; all clean after sanitize.
• Embedding fallback observability (#457) — stops silent fallback to TF-IDF when the OpenAI embedding endpoint fails; warns visibly instead of degrading silently. Also adds explicit warn-on-fallback log path.
• API issue filter tightening — /api/memories issue filter now requires strict metadata.type match instead of fuzzy heuristic.
• Graph spin fixes (5 commits) — vis-network physics now disables after stabilization, force-disables 1.5s after drag, re-enables on dragStart for elasticity, explicitly starts simulation on drag events, and stops endless spinning after memory_update realtime reload.
• Favorites filter (#544) — fetches the full memory set from the server instead of filtering only the loaded page.

Features

• Issues tab in web UI — inline status / category editing for memories of type: issue.

Reverts

• Reverted Block 4 of duplicate-review UI action-buttons wiring (held back pending follow-up work).

v0.2.27

What's New

Structured Document Storage

• memory_store_document — Store markdown documents as a root memory + searchable typed fragments (claims, plan items, references, risks, section chunks)
• memory_get_document — Retrieve documents by key with optional fragment type filtering
• memory_delete_document — Cascade delete a document and all its fragments
• Structure-aware parser (document.py) — splits by headings, tables, numbered lists, URL lists, and risk sections instead of blind token chunking
• Fragment integrity guards — protects document fragments from accidental delete, merge, absorb, duplicate detection, and supersession detection
• Graph UI — document fragments hidden from visualization (local + cloud); only the document root node is visible

Auto-categorization Improvements

• Auto-assign metadata.section from detected project at write time
• Normalize generic tags to project-prefixed tags automatically
• Backfill tag detection now catches subsection changes
• Derive subsections from bare tags when no prefixed tags exist
• Broader project detection with tag-based fallback and more content patterns

Bug Fixes

• Fix find_duplicate_candidates type error when crossref entries have mixed ID types
• Fix crossref computation to exclude document fragments from similarity graph
• Skip typed link entries (supersedes, extends, etc.) in duplicate candidate scanning

v0.2.26

What's New

memory_detect_supersessions — Retroactive supersession detection

New MCP tool that scans existing memories for pairs where one updates/replaces another, then creates supersedes edges between them. Complements memory_absorb which only catches supersessions at write time.

• Neutral LLM classification with 6-way relation enum (a_supersedes_b, b_supersedes_a, duplicate, related, contradicts, neither)
• Conservative defaults: dry_run=True, min_confidence=0.75
• 30s rate limit cooldown
• No schema changes required

Bug fixes

• Fix three bugs in memory_absorb that broke LLM classification and fact grouping
• Consolidate related new facts into single richer memories via LLM synthesis
• Bump graph cluster thresholds back to 0.4

Other improvements

• Lineage-aware retrieval: follow parameter for supersession chain walking
• Context-efficient retrieval: compact defaults, fields projection, hybrid filter fix
• Add --version flag to memora-server CLI
• Add install script for server, skill, and MCP config guidance

v0.2.25 — D1 latency collapse

Headline

`memory_create` / `memory_update` drop from 10s+ to ~2s per call on the D1 backend — 5× faster.

Three independent fixes, all landing together:

• Cached `ensure_schema()` per backend instance — previously ran 7–9 D1 round-trips on every MCP tool call (~4–8s wasted each). Now runs once per backend per process. [`memora/schema.py`]
• Single paginated LEFT JOIN for crossref/vector scan — replaces the old `list_memories(...) + _get_embeddings_for_ids(...)` two-step that cost ~10 sequential D1 round-trips on stores with many memories. [`memora/storage.py`]
• Per-instance D1 session token + backend-level keep-max bookmark mirror — the old class-level token could be stomped by cross-thread code (`cloud_sync`'s `threading.Timer`), corrupting read-your-writes. Now safe under concurrency. [`memora/backends.py`]

Also in this release

• Durable Object request reduction — WebSocket keepalive ping removed, broadcasts skipped when no clients connected, disconnect on hidden tab. Direct Cloudflare cost savings.
• Schema cache correctness — `CloudSQLiteBackend.sync_before_use()` can `shutil.move` a fresh DB file over `cache_path`; the cache now invalidates via inode-based signature so long-lived processes don't run against a post-replacement DB with stale schema assumptions.
• Security (merged from main) — XSS attribute injection fix in graph UI (#36), stale draft cancellation (#37).
• Docs — eventual-consistency model of the `related` graph documented in `memory_related` and `memory_rebuild_crossrefs` MCP tool docstrings.

Upgrade notes

• No schema migration required.
• No public API signature changes.
• No configuration changes required.
• New optional env var `MEMORA_VECTOR_SCAN_PAGE_SIZE` (default: 1000) — rarely needs tuning.
• 39/39 tests passing.

Verified performance

Measured against live D1 (steady-state, medium-length memory):

Operation	Before	After
`storage.connect()` 1st	—	~1.9s
`storage.connect()` 2nd+	~4–8s (schema probe)	~0ms (cache hit)
`add_memory`	10s+	~1.8s
`update_memory`	10s+	~1.1s

v0.2.24 — Security Hardening

Security Hardening Release

Comprehensive security audit and fixes across critical, high, and medium severity findings.

R2 Proxy Lockdown

• images/ prefix restriction on both cloud and local proxies
• Path traversal (..) blocking
• Image extension allowlist + Content-Type validation from R2 metadata

MCP Tool Rate Limiting

• Single-flight + cooldown on 7 expensive operations (rebuild, export, import, insights, duplicates, migrate)
• Prevents resource exhaustion via repeated calls

Prompt Injection Mitigation

• Memory content moved from system prompt to user/context message across all 4 LLM callsites
• Defensive markers on all memory-to-LLM paths

Upload Hardening

• Symlink chain rejection, Pillow content verification
• MIME derived from image content (not filename), no path disclosure in responses
• Extension allowlist aligned with storage backend

Additional Fixes

• Origin validation (urlparse-based) on SSE and chat endpoints
• Chat rate limiting: 30 req/min per IP
• Secure dedup cache: ~/.cache/memora/, SHA-256 filenames, fcntl locking
• SQL hardening: alias-aware ORDER BY whitelist, LIMIT/OFFSET clamping
• Error sanitization: generic messages for infrastructure errors
• Dependencies: Pillow >= 10.4.0, openai >= 1.6.0, zod 3.24.2
• Lock files: requirements.lock, requirements-local.lock, requirements-dev.lock
• Security Model documentation in README

Codebase Health (from v0.2.23)

• Full XSS sweep in graph UI (escapeHtmlAttr, CSS.escape, data-* delegation)
• PATCH metadata shallow merge with null-delete semantics
• FTS/embedding reindex on tag/metadata changes
• Module extraction: embeddings.py, schema.py split from storage.py
• Duplicate detection aligned to 0.85 threshold across all paths
• 39 tests passing

Full Changelog: v0.2.23...v0.2.24

v0.2.23

Changes

• Chat embeddings: Compute embeddings on memory create/update via chat UI, ensuring new memories appear in semantic search immediately
• Robust search: Run semantic + keyword search in parallel with isolated error handling; recent memories fallback preserved
• UI improvements: Pencil edit icon, hide complex metadata section by default
• Sync: sync-to-d1.py now syncs memories_embeddings table with transaction wrapping
• Query rewriting: Add RAG query rewriting for better retrieval

v0.2.22

What's Changed

• Fix focused node centering — nodes no longer appear shifted left when the detail panel opens
• Memora favicon — tab icon now uses the memora logo
• Wider detail panel — default width increased from 450px to 560px
• Scrubbed private data — removed hardcoded Cloudflare DB IDs from example scripts