What shipped in Remnic, release by release. Subscribe with the RSS feed.
remnic bench security injection-suite runs the H5 injection suite with per-row checkpoints, multi-host claim leases, host-fault pause, and local, Ollama, or OpenAI-compatible executors (#1962). The live executor’s default model is now qwen3.8-27b-64k:latest with a 300-second request timeout. The command reports suite status. It does not start the live factorial.
The Claude Code marketplace install now registers a working remnic MCP server at install time (#2314). It works through the plugin manifest’s userConfig and inline mcpServers. Claude Code prompts for a daemon token and routes it to the OS keychain on macOS. A stdio proxy forwards JSON-RPC to your daemon URL. No more hand-editing the bundled .mcp.json.
Contradiction detection now localizes candidates by entityRef and category before namespace-scoped search, with bounded deterministic merging (#2327). Graph-expanded recall can optionally demote paths that use stale or non-active intermediate memories (#2328). The opt-in graphPathScoring block records edge and path evidence without changing the disabled default.
The Claude Code and Codex hook health probes now send the configured bearer token. A daemon with REMNIC_AUTH_TOKEN set is detected as healthy instead of being reported as “daemon not running”, which had silently skipped auto-recall and auto-observe. Token resolution orders canonical names (REMNIC_AUTH_TOKEN) before pre-rename aliases. A stale legacy export can no longer shadow the credential the daemon runs with.
Claude Code hooks honor REMNIC_NAMESPACE / ENGRAM_NAMESPACE to target a memory namespace on recall and observe. Opt-in: unset preserves existing behavior.
Remnic Relay is a focused view for the moment two agents disagree. It shows the stale belief with its source and timestamp, routes one human-approved correction, and confirms that a cold agent session picks up the replacement and passes the same test. The mission contract and receipt API landed first, so every claim on screen traces back to a fixture event, a captured model output, a memory record, or an executed test rather than decorative UI state.
Relay pairs with MemCorrect, the memory benchmark harness that scores whether an agent accepts a correction and stops serving the stale fact afterward. One local command starts a small MCP server, runs a correction case, and writes a report, with no API key or dataset required. See the Remnic Relay design notes and the MemCorrect benchmark docs.
Coding-graph indexing used to produce nodes but zero edges, which left trace_path, blast-radius, and dead-code detection degenerate. Indexing now derives CALLS edges from parsed call sites, resolving each call to its innermost enclosing symbol and a unique same-file or import-bound target. Stale-edge deletion is provenance-scoped, so re-derivation never removes trace or LSP edges.
Separately, the example project tags in the MCP tool descriptions and test fixtures now use generic placeholders like acme-webshop and globex-storefront instead of real client names. See the coding-graph docs.
MemCorrect now routes Remnic corrections through the Correction Contract, using the public access-service surface with a plan step and a confirmed apply step. When the planner produces no applicable action, an explicit turn-path fallback takes over so a correction case never silently stalls. You select the adapter from the command line with remnic bench run memcorrect-v1 --memcorrect-adapter <remnic|prompt-only>. See the MemCorrect benchmark docs.
Recall introspection now runs through the typed Orchestrator.recallIntrospection coordinator, and the duplicate intent, QMD, and graph-recall forwarding methods are gone from the composition root. Scope-profile and team configuration parsing moved into a dedicated namespace config module, while the existing parseConfig validation and defaults stay unchanged. The result is a smaller, clearer path for asking why a memory surfaced. See the retrieval explanation docs.
Health responses now include structured QMD diagnostics: qmd.active, qmd.mode, version support, collection state, and degraded fallback status. Docker and standalone deployments can read whether QMD is really active instead of inferring it from recall behavior. That makes a misconfigured or missing search backend visible at a glance. See the search backends docs.
ensureCollection now auto-creates the QMD collection when it is missing, rather than falling back to the no-op search backend. This runs per namespace, so each namespace gets its collection created automatically on first probe. The Docker image now installs @tobilu/qmd@2.5.3 globally, so semantic search works out of the box without a manual setup step. See the search backends docs.
Coding agent mode auto-scopes memory to the current git project using an origin-URL hash, with optional branch scoping controlled by codingMode.projectScope and codingMode.branchScope. Branch scope falls back to project-level reads, so project memories stay visible from any branch while branch writes do not leak across branches. Claude Code and Codex CLI session-start hooks detect git context with a short timeout, and a review-context recall tier activates on keywords like review or what changed to boost memories that touch the modified files. See the coding agent docs.
Remnic can now import existing memory from ChatGPT, Claude, Gemini, and Mem0. Each importer ships as an optional package, loaded on demand so the base install stays small; a missing adapter prints a clean install hint instead of a module error. Run remnic import --adapter <source> --file <path> with --dry-run, --rate-limit, and --batch-size flags, or point --all-from-bundle <dir> at a folder to auto-detect every export format at once. See the importers docs.
Procedural memory now ships enabled. Fresh installs and any config that omits procedural.enabled get procedure extraction, task-initiation recall injection, and trajectory mining, using the safer defaults of a 0.75 success floor, a 14-day lookback, and at most two recalled procedures. Operators who want the previous behavior can set procedural.enabled: false explicitly. See the procedural memory docs.
Recall X-ray records exactly why each memory surfaced. The RecallXraySnapshot schema captures the retrieval tier, score decomposition, filter ladder, graph path, and audit entry ID for every result. You can read it as JSON, text, or markdown from the remnic xray "<query>" command, the GET /engram/v1/recall/xray route, or the remnic.recall_xray MCP tool. See the X-ray docs.
The LongMemEval-S and LoCoMo-10 runbook is now published, with a summary page and a benchmarks page on the site. Placeholder artifacts prove the BenchmarkArtifact v1 pipeline end to end on a fresh clone, so the reproducibility path works before any tagged numbers land. This gives a documented, repeatable way to measure recall quality. See the benchmarks runbook.
Entity files now use a two-layer format: a mutable ## Synthesis section for the current compiled truth and an append-only ## Timeline for history, tracked by synthesis_updated_at and synthesis_version metadata. Recall and briefing surfaces prefer the synthesis by default and reserve timeline snippets for history-oriented questions. Nightly governance refreshes stale synthesis in bounded batches, and remnic openclaw entities-migrate converts legacy flat entity files into the new format. See the memory lifecycle docs.
remnic openclaw install configures OpenClaw to use Remnic as its memory provider, writing the plugin entry and memory slot into ~/.openclaw/openclaw.json. It supports --yes, --force, --dry-run, --memory-dir, and --config flags, and it detects legacy openclaw-engram entries and offers migration. Remnic was formerly named engram, so older openclaw-engram configs are recognized and upgraded; see the rename notes. Expanded remnic doctor checks then verify the config file, plugin entry, memory slot, and memory directory. See the OpenClaw plugin docs.