Skip to content

v0.20.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 08 Oct 06:35
· 373 commits to develop since this release
v0.20.0
9c55dbc
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's New in v0.20.0

Minor release. Theme: one write policy for the vector index, atomic documents,
shared config resolution, and a security/robustness sweep — every index
mutation now goes through a single module, document upserts are
embed-then-commit, uteke.toml is resolved by one implementation for the CLI and
the server, GET /graph finally serves the data remember writes, and the
validated findings of the v0.19.1 review (#1319) are fixed. 29 PRs since v0.19.1.

⚠️ Behavior Change

  • List limit is capped at 1000 on /list, /recent, /contradictions, /doc/list and deprecated listings (#1321); /extract max_facts is capped at 100. Defaults are unchanged.
  • CORS is off by default (a wildcard * is opt-in), 500 responses carry a generic body, and request query strings are no longer logged (#1321).
  • Project-local .uteke/uteke.toml can no longer set credentials, endpoints or [server] (embedding.{backend,api_key,base_url,endpoint_path}, embed_fallback.* and extraction.* credentials/endpoints, the whole [server] section); set them in the global config, or UTEKE_TRUST_PROJECT_CONFIG=1 to trust the file (#1321, #1354). The server now applies the same rule and merges project sections per key instead of replacing whole sections.
  • GET /graph node ids are memory ids (they were graph-node ids), edges are the live memory_edges plus explicit graph edges, and the response gains node_id/limit query parameters and a truncated flag; edges also carry source/target (#1366, #1370). POST /graph/edge still returns the node ids it created and both id kinds are accepted.
  • Document upsert/update is atomic: if embedding fails (or no embedder is available) nothing is written, instead of leaving a half-written document (#1350). Re-parenting through doc upsert now moves the whole subtree and rejects cycles and results deeper than 10 levels (#1378).
  • Bulk "cold" forget skips pinned memories (#1333).
  • List/search/list_at_time payloads carry the stored slug, source, source_type, author_type, deprecated_at instead of defaults (#1346, #1333). Shape unchanged.
  • Auto-created similar_to / possible_duplicate links no longer point at deprecated memories and small namespaces inside a large index now get their own links (#1371). The default fusion recall does not read edges; LongMemEval fast50 (default strategy) is unchanged: recall@5 0.980, recall@10 1.000.

Security

  • Host header validation blocks DNS rebinding (#1326, #1360).
  • /import size limit, /mcp body limit and CORS, and resolve_id hardened (#1328, #1359): the import cap and its 413 message agree with the body limit, an oversized /mcp body returns 413, and ids accept only hex digits and dashes (no LIKE wildcards in the prefix scan).
  • A lock timeout no longer deletes the index files of another process; the server releases its request slot on panic; the CLI sends UTEKE_AUTH_TOKEN to the local server (#1320).
  • LIKE wildcards in document paths are escaped (#1378): _ and % in a stored path used to match unrelated documents, so moving or deleting /my_doc/ could rewrite or remove /myxdoc/....

Added

  • full_ids option for MCP uteke_list, uteke_recall and uteke_room_memories (#1357, #1358): prints full UUIDs instead of the 8-character prefix, which is ambiguous for UUIDv7 ids written within ~65 s.
  • GET /graph?node_id=&limit= and an MCP uteke_graph that lists the latest edges and takes limit (#1370).
  • Vector index sidecar consistency check on load (#1325, #1362): the .keys file carries a digest and row count of the index; a mismatch (crash between the two renames, missing sidecar) rebuilds the index from SQLite.
  • doc export writes every document to --output and honours the namespace; JSON export is an array of full documents (#1376).
  • Namespace/room/tag conventions in the bundled skill (#1356) and a refreshed skill with a command-coverage test (#1329).
  • Configuration docs: what a project config may not set.

Changed

  • index_sync: one module owns every vector index write (#1340, #1347, #1348): a single retry/persist policy, deterministic dimension errors are not retried, and an outcome that says what actually landed. 60 hand-written call sites across 7 modules went away.
  • Index saves are batched (#1322, #1363): a save happens after 64 mutations or 2 s, with flushes on shutdown and drop, instead of rewriting the whole index on every write.
  • vecq index is compacted when dead rows pass 25 % (#1324, #1364).
  • The embedder lock is released between document chunks (#1323, #1361).
  • remember no longer scans the namespace for dedup and auto-linking (one query over the candidates) and no longer holds the index write lock while wiring edges (#1368, #1371).
  • uteke.toml is resolved by one module for the CLI and the server (#1342, #1354): per-key merge, one untrusted-key policy, per-layer validation; adding a config key now needs only a struct field.
  • All memory SELECTs share one column list with the row decoder (#1339, #1346).
  • Edges to deprecated memories no longer count in graph rerank or edge traversal (#1367, #1373).

Fixed

  • FTS5 results had shifted metadata columns (#1333): slug, source, source_type, author_type, deprecated_at were misread on hybrid recall and recall_explain.
  • The exclusive index lock was lost after the first save on Unix (#1333): the flock stayed on the replaced inode.
  • doc_search paired chunk scores with the wrong rows when a chunk row was missing (#1333); resolve_db_path no longer chmods a directory it did not create (#1333).
  • Changing a document title failed with "database corruption" or left stale tokens in the search index (#1349, #1351): the documents_fts update/delete triggers now use the FTS5 'delete' command; existing stores are healed and the index rebuilt on open.
  • MCP uteke_graph_add_edge and dream contradiction edges violated the graph FK (#1365, #1370).
  • Documents: has_children stays in sync on delete and re-parent, move of a legacy row with an empty path no longer rewrites every document (#1375); a re-parenting upsert carries its descendants (#1377, #1378).
  • tags rename onto an existing tag duplicated it, VectorIndex::build on an empty store reset the dimension to 768, and recompute_importance loaded every embedding and committed per row (#1375).
  • CLI: verify-checksums --json exits 1 on a mismatch and matches the file name exactly; [lifecycle], [aging] max_access_count, update_check and doctor_footer in config files are honoured; non-ASCII titles no longer panic; doc move/update --json is valid JSON (#1334, #1376).
  • CLI recall flags the server cannot honour fall back to the local store instead of being dropped; list --entity/--category paginates the matches; doc list --tree prints descendants; upgrade verifies every new binary before installing any; the batch-import bail-out and onboard TOML escaping and legacy migrate_content are fixed (#1376).
  • uteke upgrade verification no longer fails with "Text file busy" on a freshly written binary (#1338).

Docs

  • Benchmark tooling: make_fast_eval.py finds its id list, modal_fanout.py takes LMEVAL_MODEL_DIR/UTEKE_RELEASE, dataset size corrected (#1374).

Dependencies

  • uuid 1.27.0, usearch 2.26.3.

📦 Downloads

Each archive contains three binaries + ONNX Runtime shared library:

Binary Description
uteke CLI tool
uteke-serve HTTP server daemon
uteke-mcp MCP server (stdio + HTTP)
libonnxruntime.so* / libonnxruntime*.dylib / onnxruntime*.dll ONNX Runtime
Platform File
Linux x86_64 (AVX2) uteke-x86_64-unknown-linux-gnu-v0.20.0.tar.gz
Linux x86_64 (Legacy, SSE4.2) uteke-x86_64-unknown-linux-gnu-legacy-v0.20.0.tar.gz
Linux ARM64 uteke-aarch64-unknown-linux-gnu-v0.20.0.tar.gz
macOS Apple Silicon uteke-aarch64-apple-darwin-v0.20.0.tar.gz
Windows x86_64 uteke-x86_64-pc-windows-msvc-v0.20.0.zip

Legacy Bundle (Linux only) — Includes a SSE4.2-only ORT sidecar (ort-legacy/) for CPUs without AVX2 (Intel Celeron J4125/N4020). Use this bundle to avoid SIGILL crashes on older hardware.

🚀 Quick Start

# Quick install (Linux / macOS)
curl -fsSL https://raw.githubusercontent.com/codecoradev/uteke/main/install.sh | sh

# Pin a specific version
UTEKE_VERSION=v0.20.0 curl -fsSL https://raw.githubusercontent.com/codecoradev/uteke/main/install.sh | sh

# Store a memory
uteke remember "Important context" --tags project

# Recall by meaning
uteke recall "what was that context?"

# Start server for fast AI agent access
uteke-serve --port 8767

Full changelog: https://github.com/codecoradev/uteke/blob/main/CHANGELOG.md

Community contributors

(No external contributions in this cycle. Full contributor list: https://github.com/codecoradev/uteke/graphs/contributors)

Assets 8
Loading