memories.sh logomemories.sh

Changelog

Product and documentation updates.

February 28, 2026

  • Launched segmented memory lifecycle messaging across docs and homepage:
    • Session memory (working context)
    • Semantic memory (durable truths/preferences)
    • Episodic memory (daily logs + raw snapshots)
    • Procedural memory (reusable workflows)
  • Added new Concepts guide: Memory Segmentation, including trigger patterns, compaction guidance, and storage routing.
  • Updated homepage copy to foreground lifecycle segmentation, compaction-safe checkpointing, and deterministic OpenClaw memory flow.
  • Updated Getting Started + README guidance with explicit lifecycle commands:
    • memories session start|checkpoint|snapshot|end
    • memories compact run
    • memories openclaw memory bootstrap|flush|snapshot|sync
  • Expanded integration docs with lifecycle-first operational patterns:
    • SDK: budget/session hints + checkpoint-first flow
    • MCP: lifecycle sequence for long-running sessions
    • OpenClaw: deterministic bootstrap → flush → snapshot → sync routine

February 26, 2026

  • Added memory lifecycle observability endpoint: GET /api/sdk/v1/management/memory/observability (lifecycle, compaction, consolidation, contradiction trend metrics + alarms).
  • Added replay/eval harness endpoint: POST /api/sdk/v1/management/memory/eval/replay for extraction F1, compaction retention, and trigger-match scoring.
  • Added default-on lifecycle feature flags with explicit endpoint gating:
    • MEMORY_SESSION_ENABLED (sessions API surface)
    • MEMORY_COMPACTION_ENABLED (memory lifecycle observability/eval)
    • MEMORY_CONSOLIDATION_ENABLED (memories/consolidate)
    • MEMORY_PROCEDURAL_ENABLED (skills/files/promote)
  • Added typed disabled responses for lifecycle flags (MEMORY_SESSION_DISABLED, MEMORY_COMPACTION_DISABLED, MEMORY_CONSOLIDATION_DISABLED, MEMORY_PROCEDURAL_DISABLED).
  • Deprecated legacy strategy aliases in @memories.sh/core (baseline, hybrid_graph) with one-time warnings and suppression support via MEMORIES_SUPPRESS_DEPRECATION_WARNINGS=true.
  • Updated SDK docs/contract matrix with session, consolidate/promote, and memory management observability/eval routes plus Phase 7 rollout defaults.

February 21, 2026

  • Added cron reminder support in CLI via memories reminders (add, list, run, enable, disable, delete).
  • Added local CLI MCP reminder tools: add_reminder, list_reminders, run_due_reminders, enable_reminder, disable_reminder, delete_reminder.
  • Published reminder docs in CLI reference and MCP references, including local-only reminder tooling guidance.
  • Added first-class memory graph nodes (node_type = memory, node_key = memory_id) with self links so each memory is directly addressable in graph traversal.
  • Updated graph mapping cleanup to remove edges attached to deleted memory nodes, not only edges where the memory is listed as evidence.
  • Extended vacuum_memories cleanup to prune graph mappings for purged soft-deleted memories.
  • Added embedding-driven similar_to graph edges between memory nodes (bidirectional) using cosine thresholding and per-memory caps.
  • Added similarity edge controls: SIMILARITY_EDGE_THRESHOLD, SIMILARITY_EDGE_MAX_K, SIMILARITY_EDGE_MAX_PER_MEMORY.
  • Added optional LLM relationship extraction for ambiguous similarity matches to create contradicts and supersedes memory edges.
  • Added graph LLM controls: GRAPH_LLM_AMBIGUOUS_SIMILARITY_MIN, GRAPH_LLM_AMBIGUOUS_SIMILARITY_MAX, GRAPH_LLM_RELATIONSHIP_CONFIDENCE_THRESHOLD, GRAPH_LLM_RELATIONSHIP_MODEL_ID.
  • Added relationship-aware graph ranking weights by edge type (caused_by, contradicts, supersedes, similar_to, etc.) with confidence + hop decay in retrieval scoring.
  • Added conflicts to get_context responses when returned memories are linked by contradicts edges, including confidence and clarification suggestions.
  • Added semantic LLM relationship extraction on writes for caused_by, prefers_over, depends_on, specializes, and conditional_on edges (with reusable condition graph nodes).
  • Made relationship edge replacement atomic via savepoint rollback in graph upsert (replaceMemoryRelationshipEdges) to prevent partial edge loss on write failures.
  • Added structured LLM extraction degradation reporting in embedding worker metrics (GRAPH_RELATIONSHIP_PARTIAL_DEGRADE) with issue-level codes (GRAPH_LLM_CLASSIFICATION_FAILED, GRAPH_LLM_SEMANTIC_EXTRACTION_FAILED).
  • Clarified embedding observability/operations docs so graph extraction degradation is mapped to worker.topErrorCodes with structured issue summary format.
  • Optimized graph-only retry path after GRAPH_RELATIONSHIP_SYNC_FAILED to reuse stored embeddings and avoid redundant embeddings API calls.
  • Added live Graph Explorer status refresh (manual + background polling) so rollout health, alarms, and quality gate data stay current without page reloads.
  • Added keyboard-accessible graph canvas interactions in Graph Explorer (focusable node/edge controls with Enter/Space activation and ARIA labels).
  • Updated Graph Explorer node switching to cancel stale /api/graph/explore requests using AbortController, reducing redundant in-flight fetches during rapid navigation.
  • Stabilized Graph Explorer wheel zoom behavior so rapid scroll events compose against the latest viewport state without dropping zoom steps.
  • Clamped Graph Explorer mini-map viewport geometry at low zoom to keep viewport bounds inside the graph canvas extent.
  • Added Graph Explorer interaction logic tests (filtering, selected-edge fallback, zoom/minimap math, and node-stat summaries).
  • Persisted Graph Explorer filters in URL state (edge/node type selections, thresholds, evidence-only) for shareable deep links.
  • Added multi-key API key management: users can create multiple active mem_ keys, list key metadata, and revoke individual keys without rotating all keys.
  • Updated memory graph architecture docs to document memory nodes and self-link behavior.
  • Expanded graph + MCP docs with edge weight tables, contradiction/semantic extraction pipeline guidance, conflict payload schemas/examples, and skills guidance for conflict-aware agent handling.

February 20, 2026

  • Removed legacy tenant management routes /api/mcp/tenants and /api/sdk/v1/management/tenants.
  • Removed legacy tenant compatibility wrapper code and telemetry write-paths now that traffic is zero.
  • Updated tenant-routing and SDK endpoint contract docs to reference only /api/sdk/v1/management/tenant-overrides.
  • Enforced hashed-only CLI bearer token auth by removing plaintext cli_token runtime fallback in authenticateRequest.
  • Added migration 20260220053000_backfill_cli_token_hashes.sql to backfill cli_token_hash from legacy cli_token and clear plaintext values.
  • Upgraded embedding billing token accounting to prefer provider/model tokenizer counts (OpenAI embeddings) before fallback heuristics.
  • Added embedding token discrepancy metrics (token_count_method, input_tokens_char_estimate, input_tokens_delta) via migration 20260220062000_add_embedding_token_accounting_columns.sql.
  • Added hosted SDK embeddings docs at /docs/sdk/embeddings with write-path model selection, retrieval strategy controls, and migration guidance.
  • Added embeddings endpoint contract documentation for:
    • GET /api/sdk/v1/embeddings/models
    • GET /api/sdk/v1/management/embeddings/usage
    • GET|POST /api/sdk/v1/management/embeddings/backfill
    • GET /api/sdk/v1/management/embeddings/observability
  • Documented retrieval strategy defaults for context/get and memories/search as lexical, with compatibility aliases:
    • baseline -> lexical
    • hybrid_graph -> hybrid
  • Documented embeddingModel request support on memories/add and memories/edit, including response metadata (embeddingModel, embeddingModelSource).
  • Clarified retrieval trace diagnostics and fallback semantics to support phased semantic/hybrid rollout.
  • Set graph retrieval feature flag default to enabled (GRAPH_RETRIEVAL_ENABLED=true fallback) and added rollout decision payloads with SLO/gap metrics (rolloutPlan) on GET|POST|PATCH /api/sdk/v1/graph/rollout.
  • Added graph default-on rollout policy docs with explicit go/no-go thresholds, baseline gap reporting, and GRAPH_ROLLOUT_AUTOPILOT_ENABLED behavior.
  • Added default retrieval strategy policy autopilot (GRAPH_DEFAULT_STRATEGY_AUTOPILOT_ENABLED, default true) with two-window promotion (lexical -> hybrid) and automatic rollback on gate regression.
  • Updated context/get and graph/trace to apply policy defaults when strategy is omitted and return retrieval-policy trace diagnostics.
  • Added retrievalPolicy snapshot payload to GET|POST|PATCH /api/sdk/v1/graph/rollout.
  • Added scheduled Graph Rollout Baseline Report automation (every 6 hours) that persists reports/graph-rollout/{history,latest}.json + latest.md with blocker trends, recommendation history, and readiness regression detection.
  • Added memories setup --minimal-local preset for local-only onboarding with no cloud/workspace dependency.
  • Added memories doctor --local-only to skip cloud checks and provide cleaner local-mode diagnostics.
  • Documented a 10-minute local happy path (setup -> doctor --local-only -> add -> search) across CLI and Getting Started docs.
  • Added Local Onboarding Smoke GitHub Actions workflow to validate clean-machine minimal-local onboarding on CLI-related PRs.

February 16, 2026

  • Deprecated legacy tenant management routes: /api/mcp/tenants and /api/sdk/v1/management/tenants.
  • Moved both legacy tenant paths to compatibility wrappers over /api/sdk/v1/management/tenant-overrides.
  • Added consistent deprecation headers (Deprecation, Sunset, Link, Warning) for legacy tenant route responses.
  • Added server-side legacy route telemetry table (legacy_route_usage_events) to track migration progress to zero traffic before removal.
  • Published sunset timeline and migration guide at /docs/sdk/legacy-tenant-route-sunset.

February 14, 2026

  • Released @memories.sh/cli v0.7.6 and @memories.sh/core v0.3.2.
  • Fixed older_than_days float handling — values like 0.5 now round up to 1 instead of silently flooring to 0.
  • Replaced O(n) graph mapping cleanup with batched IN-clause deletes for bulk_forget_memories.
  • Optimized bulk_forget_memories dry run to use a single bounded query instead of unbounded COUNT(*).
  • Empty filter arrays (types: [], tags: []) are now correctly treated as no filter instead of bypassing validation.
  • Improved error messages to clarify that project_id alone is not a sufficient filter for bulk operations.
  • Updated tags documentation to warn about substring matching (e.g., "test" also matches "testing").
  • Updated pattern documentation to clarify glob syntax (* and ?) instead of SQL LIKE.

February 13, 2026

  • Added bulk_forget_memories tool and POST /api/sdk/v1/memories/bulk-forget endpoint for filtered mass soft-deletion of memories by type, tags, age, content pattern, or project. Supports dry_run mode for previewing matches before deleting.
  • Added vacuum_memories tool and POST /api/sdk/v1/memories/vacuum endpoint for permanently purging all soft-deleted records to reclaim storage space.
  • Added client.memories.bulkForget() and client.memories.vacuum() methods to @memories.sh/core.
  • Added bulkForgetMemories() and vacuumMemories() tool wrappers to @memories.sh/ai-sdk.
  • MCP server now exposes 9 tools (up from 7).

February 12, 2026

  • Added workspace-level graph rollout controls with off, shadow, and canary modes.
  • Added shadow rollout metrics and fallback safety alarms (HIGH_FALLBACK_RATE, GRAPH_EXPANSION_ERRORS) to graph status reporting.
  • Added retrieval quality evals (current vs previous window) and hard canary gate enforcement (CANARY_ROLLOUT_BLOCKED) when fallback/relevance metrics regress.
  • Added organization audit logs/event trail with storage migration, /api/orgs/[orgId]/audit endpoint, and Team dashboard audit panel.
  • Added SDK HTTP rollout endpoint docs and hybrid retrieval trace guidance, including applied strategy + fallback diagnostics.
  • Added dedicated SDK graph documentation: architecture overview and graph operations runbook.
  • Upgraded dashboard graph explorer with focus mode, filter controls, fit view, mini-map, and deep-linkable node/edge state.
  • Added guided CLI setup mode selection (auto / local / cloud) and upgraded memories doctor --json with stable check codes, categories, and aggregated next-step remediation.
  • Added workspace-switch performance budgets (p50 ≤ 400ms, p95 ≤ 1000ms over 24h) with integration-health alerts and CI enforcement for budget evaluation logic.
  • Added automatic GitHub capture intake (PRs, issues, commits) with dashboard review/approval queue before writing to memory.
  • Added Actionable Intelligence dashboard panel with stale-rule detection, conflict candidates, weekly change summary, and suggested archive/merge/relabel actions.
  • Added one-click Actionable Intelligence apply flow (archive, merge, relabel) with backend execution API and immediate dashboard memory-list updates.
  • Optimized workspace switching with cached workspace-summary API responses, switcher prefetch for personal/org summaries, and batched org-member login lookups to remove per-member N+1 auth requests.
  • Expanded GitHub auto-capture depth with release-note webhook adapters (release events) and richer queue filters (event, repo, actor, workspace, q) for review workflows.

February 11, 2026

  • Updated OpenClaw integration docs to follow workspace onboarding (openclaw onboard) with generated artifacts in ~/.openclaw/workspace (AGENTS.md and skills/).
  • Clarified that OpenClaw integration is documented via workspace artifacts, not .openclaw/mcp.json.

Starter Apps Quickstart

Launch Next.js, Express, or Python starter apps and run add/search/context flows in under 10 minutes.

CLI Reference

Complete reference for all memories.sh CLI commands.

On this page

February 28, 2026February 26, 2026February 21, 2026February 20, 2026February 16, 2026February 14, 2026February 13, 2026February 12, 2026February 11, 2026