Skip to main content

What's new

Changes since the 1.10 line, newest first. Each entry links to the page that documents it in full.

Verified caller scope and tenant isolation​

The REST, A2A and MCP HTTP surface can now run behind verified credentials instead of trusting the network.

  • AUTH_MODE=jwt: every route except /health and /ready requires a signed bearer token. A new route is protected by default.
  • One database per tenant. The tenant comes only from the verified token and selects the SurrealDB namespace (server mode) or directory (embedded mode).
  • Safe default. API_HOST now defaults to 127.0.0.1, and AUTH_MODE=local is refused on any other bind address.
  • Operator permission. Retry, reject and re-key of durable operations need an explicit operator claim.
  • User-restricted tokens limit a credential to one user's memories.
  • MCP sessions (streamable HTTP and legacy SSE) are bound to the principal that opened them.

See Authentication and tenancy.

:::warning Upgrade note If you bound the server to a non-loopback address before, it will now refuse to start until you either set AUTH_MODE=jwt with the AUTH_JWT_* settings or bind API_HOST to 127.0.0.1. The Docker Compose files now require the AUTH_JWT_* variables and publish only on the host loopback. :::

Expired memory is invisible, and cleanup reports honestly​

A memory with valid_until in the past no longer appears in any normal read, even before the cleanup worker runs, and cannot be updated, renewed, merged into by a near-duplicate write, or copied into a summary. The cleanup worker reports what it actually removed, and a failed deletion is counted, logged and retried. See Memory expiry.

Evidence-backed compression​

compress_memories and auto_summarize_task_stream now publish summaries whose claims quote their sources. A deterministic gate checks every quote and the token budget; the sources are archived, not deleted, and stay readable as evidence. The default summarizer is extractive and never leaves the process. See Compression.

Hybrid search works as documented​

The BM25 leg of hybrid search used a full-text index that did not exist and swallowed the error, so hybrid search was silently vector-only. Migration v24 creates the index. See Migrations.

SurrealDB 4.0 and SurrealKV​

The server targets SurrealDB 4.0.0. Embedded mode now runs on SurrealKV, because 4.0 removed RocksDB. A RocksDB data directory is converted in place on the first 4.0 start, which is one-way, so back it up first. See Storage and Deployment.

Durable operations: retry, reject and stats​

POST /api/v2/operations/{id}/retry, POST /api/v2/operations/{id}/reject and GET /api/v2/operations/stats give operators control over stuck work, and operations that can never succeed are rejected instead of being retried forever. See Durable operations.

Claude Code hooks and statusline​

surreal-memory-server hook session-start|pre-compact|session-end and surreal-memory-server statusline recall and persist session memory through the running server. They never block the harness. See CLI and hooks.

Query embedding cache​

Search keeps a bounded cache of query embeddings, keyed by the provider and model identity so two models can never share an entry. Its hit and miss counts are in GET /api/v2/operations/stats. Capacity is set with SURREAL_MEMORY_QUERY_EMBED_CACHE. See Configuration.

Migrations v22 to v24​

VersionNamePurpose
22compression_derivations_and_claimsSummary derivations and their claims
23compression_evidence_and_memory_statusEvidence rows and archived-source status
24memory_fulltext_indexThe BM25 index hybrid search relies on

All three are additive and apply automatically on startup.