What's Changed

• docs: Ollama local-first setup + pipeline stage count alignment by @rudi193-cmd in #9
• fix(security): escape LIKE wildcards in namespace path resolution by @rudi193-cmd in #10

Full Changelog: v0.2.10...v0.2.11

What's Changed

• docs: post-install guide + fix MCP copy buttons (#6) by @rudi193-cmd in #8
• fix: pattern stage silently drops all output due to && vs || in fallback logic by @ali-h-abbas in #7
• docs(brain): align Consolidate docstring with the 8-stage pipeline by @truffle-dev in #1

New Contributors

• @rudi193-cmd made their first contribution in #8
• @ali-h-abbas made their first contribution in #7
• @truffle-dev made their first contribution in #1

Full Changelog: v0.2.9...v0.2.10

What Changed

Consolidation Quality Fixes

• Removed ungrounded word validation — the old check rejected valid generalized facts. Generalization IS the point of consolidation; the system prompt already constrains the LLM, and downstream dedup/contradiction/decay catch real problems.
• Improved reasoner prompt — now instructs the LLM to preserve specifics (proper nouns, library names, version numbers, file paths, quantities) instead of producing vague summaries.
• Lowered dedup threshold from 0.95 → 0.85 — aligns with the similarity threshold used for clustering; the old 0.95 let near-identical facts through.
• Boosted confidence for structured facts — facts with entity+property+value fields get a 30% confidence boost since structured extraction requires more LLM precision.

Documentation

• Added MCP client setup section to README — SSE URL (http://localhost:8080/sse) plus example configs for Cursor, Claude Desktop, OpenCode, and Windsurf.

Results

• Fact creation rate went from 1/24 to 5/5 in testing
• Zero grounding rejections on valid facts
• Specifics (numbers, names, thresholds) now preserved in fact content
• Structured facts at confidence 0.53 vs old 0.33

Summary

• Scopes MCP context tools to explicit namespaces instead of defaulting to /.
• Tunes MCP prompts so context is optional short-lived working focus, while recall remains the primary durable-memory lookup path.
• Updates runtime version metadata to 0.2.8.

Verification

• go test ./...

What's New

Prompt & Consolidation Fixes

• Reasoner: Changed → with strict anti-verbatim rules
  • First-person stripping (// removal)
  • Explicit bad/good JSON examples in prompt
  • Grounding validation now rejects substring copies and >80% word overlap with any source episode

MCP Prompt Enhancements

• Server Description: Added Proactivity Clause + Tool Decision Tree for fast tool selection
• remember: Added Trash Filter — session narration, unverified hunches, temporary states, and generic platitudes are now explicitly discouraged. Added gate.
• consolidate: Clarified that and return EMPTY for unconsolidated namespaces — making clear why consolidation is non-optional.
• create_goal: Added Quality Gate — one-off tasks finishing in a single conversation should not become goals.

Version

MCP server version: →

Feature

Added

• Full consolidation result returned from the consolidate MCP tool
• Previously: only 4 fields (namespace, facts_created, relationships_found, patterns_found)
• Now: all 20 fields from ConsolidationResult

New Fields in Response

• episodes_read - number of episodes processed
• facts_deduplicated - duplicate facts skipped
• causal_links_found - causal relationships extracted
• contradictions_found / contradictions_auto_resolved
• goals_annotated / goals_suggested_complete
• failure_repeats_detected / failure_patterns_found
• hypotheses_auto_confirmed / hypotheses_auto_rejected / hypotheses_updated
• facts_decayed / facts_expired - confidence decay stats
• llm_calls - number of LLM API calls made
• duration - time taken (human-readable string)
• errors - list of any errors encountered

Why

Agents using the consolidate tool can now see the full picture of what happened, including errors and all processing stages.

Bug Fix

Fixed

• Consolidation now only advances checkpoints on success
  • Episodes, facts, patterns with failed extraction are NOT lost
  • If any error occurs during consolidation, the checkpoint is NOT advanced
  • Failed clusters are automatically retried on the next consolidation run
  • Commit: d4cf1e4

Technical Details

• Modified consolidateEpisodesToFacts(): Only update LastEpisodeID if no errors
• Modified consolidateFactsToRelationships(): Only update LastFactID if no errors
• Modified consolidateToPatterns(): Added success flag to checkpoint update
• Updated updatePatternCheckpoint(): Only advance LastPatternFactID/LastPatternRelID on success

Why This Matters

Before: If consolidation hit an error during extraction, the checkpoint still advanced. Next run would skip the failed episodes forever.

After: Checkpoint only advances when consolidation fully succeeds. Failed episodes are retried. Deduplication prevents duplicate facts from already-succeeded clusters.

Result: Consolidation is now fully idempotent and safe from data loss. Episodes/facts/patterns can never be silently dropped.

Feature

Added

• Idempotent namespace creation: CreateNamespace now never fails on duplicate slugs
  • If namespace exists: updates name/description and returns id
  • If namespace new: creates it as normal
  • Commit: b782dd7

Technical Details

• Added ON CONFLICT (slug) DO UPDATE clause to namespace INSERT
• Updates name and description if namespace already exists
• Also updates updated_at timestamp on update
• Consistent with parent namespace auto-creation behavior

Benefit

• Auto-initialization of /self and /self/* namespaces no longer fails on second run
• Makes the initialization process fully idempotent
• Enables safe re-runs of configuration/setup scripts

Bug Fix

Fixed

• Database type conversion error: Fixed 'unable to encode int into text' error when storing vector dimension
  • Settings table stores all values as TEXT columns
  • Dimension values now properly converted to string before INSERT
  • Dimension values properly parsed back from string when reading
  • Commit: 4893334

Technical Details

• Added strconv import to internal/db/db.go
• Use fmt.Sprintf("%d", expectedDim) to convert int to string for storage
• Use strconv.Atoi(storedDimStr) to parse string back to int when validating

This patch ensures v0.2.2's dimension-based validation actually works in practice.

Changes

Features

• Dimension-based validation: Change from model-specific validation to dimension-based validation
• Flexible embedding model switching: Users can now switch between different embedding models as long as they output the same dimension
  • Example: Switch from gemini-embedding-001 to text-embedding-3-small (both 1536D) without losing data
  • Example: Switch from any 1536D model to any other 1536D model
• Model metadata for audit: Embedding model name is still stored in settings for monitoring/audit purposes, but doesn't constrain storage

Technical Details

• Store vector_dimension in settings table (the actual storage constraint)
• Store embedding_model separately for audit purposes (flexibility constraint)
• Dimension validation via validateDimensionLock() function
• Model metadata recording via storeEmbeddingModelMetadata() function
• Improved error message: explains that users can switch models with the same dimension
• Fully backward compatible: existing databases work seamlessly

Benefits

• No vendor lock-in to specific embedding model provider
• Easy optimization: switch models for better quality/cost without database recreation
• Future-proof: new embedding models can be adopted without data loss
• Dimension is the source of truth, not the model name