Remind

Appearance

Changelog ​

All notable changes to Remind.

[Unreleased] ​

[0.10.5] - 2026-05-05 ​

Added ​

  • Dual-track concepts — Concepts are now classified as either pattern (generalizations from observations and decisions) or fact_cluster (verbatim fact details that are never abstracted away). Fact clusters include a summary generated from their constituent facts.
  • Entity embeddings — Entities are now embedded alongside episodes for richer retrieval.

Changed ​

  • Consolidation prompts — Improved concept and fact-tracking prompts for better specificity and throughput; fact clustering logic tuned for higher precision.

Removed ​

  • Plan / spec / task episode types and CLI — The spec, plan, and task episode types, related CLI commands (specs, plans, tasks, task …), and MCP tools have been removed. Remind is a memory layer, not a task planner. Use standard episode types (observation, decision, question, fact, etc.) with entities and metadata to track work. Existing databases with these episode types continue to work; the types are just no longer surfaced by the CLI or MCP.
  • Bundled plan/spec/implement skills — remind-plan, remind-spec, and remind-implement skills removed from the package. skill-install now installs only the base remind skill.

Fixed ​

  • Entity processing bug — Fixed a bug where entity extraction could produce incorrect associations during consolidation.
  • UI fixes — Various web UI reliability improvements.

Documentation ​

  • README, MCP reference, CLI reference, and concept pages updated to reflect removed types and current CLI (topics subgroup, types command, corrected ingest topic behavior, expanded MCP tool list).

[0.10.4] - 2026-04-09 ​

Added ​

  • cli_output_mode / REMIND_CLI_OUTPUT_MODE — Default table, or json / compact-json for machine-readable browse output; --json, --compact-json, and --table override per run (at most one).
  • JSON output — inspect, specs, plans, tasks, topics list/overview, entities, search, mentions, decisions, questions, types, status, entity-relations, deleted support --json / cli_output_mode=json.
  • --compact-json — Minimal id / title / summary payloads (plus per-command fields such as task plan_id/plan, search score, entities mention_count, status metadata wrapper).
  • remind tasks — Plan column; --by-plan groups by plan then status; JSON includes enriched plan and optional by_plan shape.

Documentation ​

  • Docs site — Fixed navigation links and expanded the SQLite examples page.

[0.10.3] - 2026-04-08 ​

Added ​

  • Topic reassignment — Existing episodes and concepts can be moved between topics (or cleared) via update-episode / update-concept (--topic, --clear-topic), REST topic field on updates, and MCP topic (empty string clears).

[0.10.2] - 2026-04-08 ​

Fixed ​

  • SQLite startup crash with sqlite-vec — If sqlite-vec was installed but Python’s sqlite3 could not load extensions (common on macOS / some pyenv builds), Remind could crash on first DB connection. Remind now detects this and falls back to brute-force vector similarity.

Documentation ​

  • SQLite vector search — Requirements for native sqlite-vec indexes, how to verify enable_load_extension, and typical Homebrew + pyenv setup are documented under Configuration and Retrieval, plus the SQLite example README.

[0.10.1] - 2026-04-08 ​

Added ​

  • re-embed command — Regenerate stored embeddings after changing embedding model or dimensions.

[0.10.0] - 2026-04-08 ​

Added ​

  • Native vector indexes — Retrieval now uses sqlite-vec on SQLite and pgvector on PostgreSQL when available, with automatic brute-force cosine fallback.
  • Cross-encoder reranking — Optional reranking stage for recall (sentence-transformers) with configurable recall_initial_candidates.
  • Hybrid retrieval scoring — Embedding similarity is fused with keyword overlap via configurable hybrid_keyword_weight.
  • Topics as first-class partitions — topic support on episodes/concepts with topic-aware consolidation and retrieval behavior.
  • Topic APIs/tools — list_topics and topic_overview available in CLI, MCP, and Python API.
  • Ingest controls — Optional ingest instructions plus topic inference when topic is omitted.
  • Configurable episode_types — Plan/spec/task CLI and MCP surfaces register only when enabled.
  • Project-local config — .remind/remind.config.json supported alongside project database.
  • SQLAlchemy databases — PostgreSQL and MySQL support through REMIND_DB_URL / db_url.
  • Parallel consolidation and richer stats — Faster batch consolidation and episode type breakdowns in stats.
  • Memory metadata additions — source_type on episodes and supersedes relation surfaced in recall.

Changed ​

  • Auto-ingest — Removed configurable density-threshold gating; improved chunking and text processing.
  • Consolidation & extraction — Tuned prompts, token limits, and concurrency defaults for better quality/throughput.
  • Retrieval runtime — Improved recall performance, including reranker warm-up behavior in CLI workflows.
  • Web UI — Better handling for custom episode types, topics, task flows, and non-SQLite database URLs/paths.

Fixed ​

  • Reconsolidation edge cases.
  • MCP episode argument parsing.
  • Task creation from the web UI.
  • Web UI and stats with PostgreSQL and other SQLAlchemy URLs.
  • Embedding dimension handling and reranker NaN-score safety.

[0.8.0] - 2026-03-23 ​

Added ​

  • Episode embeddings — remember now embeds episode content by default for direct vector search during recall. Use --no-embed to skip (faster, no API call).
  • Direct episode recall — recall --episode-k N (CLI) or episode_k parameter (Python/MCP) retrieves episodes by embedding similarity alongside concept-based spreading activation. Default: 5. Set to 0 to disable.
  • embed-episodes command — Backfill embeddings for episodes created before episode embedding was enabled.
  • Contradiction display — Recall output now shows inbound and outbound contradicts relations per concept, with context.
  • Batched contradiction detection — Consolidation compares new material against existing concepts in configurable batches (consolidation_concepts_per_pass, default: 64).

Changed ​

  • Azure OpenAI provider upgraded to OpenAI v1 API; api_version config removed, /openai/v1 appended to base URL automatically
  • Default episode_k set to 5 for direct episode recall

Fixed ​

  • Ingest buffer handling in foreground mode
  • Async processing fixes in background worker
  • Contradiction retrieval improvements (batched comparison against existing concepts)

[0.7.0] - 2026-03-19 ​

Added ​

  • Auto-ingest pipeline — ingest() and flush_ingest() for automatic memory curation. Buffers raw text, scores information density via LLM, and distills memory-worthy episodes automatically. Available via CLI, MCP, and Python API.
  • Fact episodes — New fact episode type for specific factual assertions (config values, names, dates, technical details). Consolidation preserves fact details verbatim rather than generalizing them away.
  • Outcome episodes — New outcome episode type for action-result pairs with structured metadata (strategy, result, prediction_error). Consolidation extracts causal strategy patterns.
  • Entity name matching in retrieval — Queries now match against entity names directly (fast, no embedding needed), complementing semantic search.
  • Minimum activation floor — Retrieval drops concepts below a configurable min_activation threshold (default: 0.15), reducing low-relevance noise.
  • Entity deduplication — Entity extraction deduplicates by name across types to prevent duplicates like family:Capulet and character:Capulet.
  • status CLI command — Shows processing status: running workers, pending episodes, queued ingest chunks.
  • Per-provider ingest model — Configure a separate (cheaper/faster) model for triage without affecting consolidation quality.
  • Debug file logging — Enable logging_enabled in config to get full LLM prompt/response logs in remind.log next to the database.
  • Background ingest worker — CLI ingest command queues work and spawns a background worker by default. Use --foreground for synchronous processing.
  • Batch consolidation progress — consolidate and reconsolidate commands show per-batch progress for large runs.
  • Hybrid recall episodes — Retrieval now returns source episodes with type labels and entity context alongside concepts.

Changed ​

  • Default recall k reduced from 5 to 3 for more focused results
  • recall no longer requires a query when --entity is provided
  • end-session now flushes the ingestion buffer before consolidating
  • Consolidation loops through all batches internally instead of requiring external batch loop
  • Foreground consolidation acquires a file lock to prevent concurrent runs

[0.6.0] - 2026-03-09 ​

Added ​

  • Task management system: task episodes with status tracking (todo, in_progress, done, blocked), dependency chains, plan/spec linking, and priority levels
  • Agent workflow skills for the plan-to-implementation lifecycle:
    • remind -- base memory operations reference
    • remind-plan -- interactive planning with sparring and crystallization
    • remind-spec -- spec-driven development with lifecycle management
    • remind-implement -- systematic task execution loop
  • Active tasks are excluded from consolidation; completed tasks become eligible

Changed ​

  • Updated web UI

Fixed ​

  • run_async helper function

[0.5.3] - 2026-03-03 ​

Fixed ​

  • When updating episodes, reset entity associations so they are rebuilt on next consolidation

[0.5.2] - 2026-03-02 ​

Fixed ​

  • Actually include web UI static files in wheel builds

[0.5.1] - 2026-03-02 ​

Fixed ​

  • Include built web UI assets in package

[0.5.0] - 2026-03-02 ​

Added ​

  • Support for updating and deleting concepts and episodes
  • Built-in memories about Remind itself for self-aware assistance

Changed ​

  • Consolidation is now non-blocking, improving CLI responsiveness

Fixed ​

  • Explicit consolidation command now works correctly
  • Build issues resolved

[0.4.0] - 2026-02-26 ​

Added ​

  • Memory decay system: concepts gradually lose retrieval priority based on recall frequency
    • decay_factor (0.0--1.0) multiplies retrieval activation score
    • Decay runs every N recalls with configurable rate
    • Rejuvenation: recalled concepts receive activation-proportional boost
    • 60-second grace window protects recently-accessed concepts
  • DecayConfig in config file under "decay" key
  • Metadata table in SQLite for persistent key-value storage
  • Memory status panel in web UI

[0.3.1] - 2026-02-25 ​

Fixed ​

  • Bump numpy requirement to >=2.0.0 to prevent segfault on macOS in sandboxed environments

[0.3.0] - 2026-02-25 ​

Added ​

  • Global config file support (~/.remind/remind.config.json)
  • --version CLI argument
  • Collapsible sidebar in web UI
  • Agent skills support for Claude Code integration

Fixed ​

  • Entity relationship extraction
  • Consolidation with explicit entities

[0.2.0] - 2026-01-09 ​

Added ​

  • Web UI with interactive concept graph visualization (D3-based)
  • Docker support
  • Entity inspection UI and MCP tools
  • LLM-powered query answering with source episodes
  • Dark mode support
  • Batch consolidation and reconsolidation

[0.1.0] - 2026-01-04 ​

Added ​

  • Core memory system with episodes and concepts
  • Spreading activation retrieval algorithm
  • LLM-powered consolidation
  • Entity extraction from episodes
  • Provider support: Anthropic, OpenAI, Azure OpenAI, Ollama
  • MCP server (SSE mode)
  • CLI tool
  • SQLite persistence layer
  • Background consolidation
  • Project-aware database paths

Released under the Apache 2.0 License.