Blog draft: From Storage to Intelligence — The v10.55→v10.57 Arc

[filhocf]

Hey @doobidoo — you mentioned in #870 that you'd review a blog post if I drafted one around the v10.55→v10.57 arc. Here it is.

Feel free to edit, cut, or redirect. Once you're happy with it, I can cross-post wherever makes sense (dev.to, GitHub blog, etc.).

---

From Storage to Intelligence: The v10.55→v10.57 Arc in mcp-memory-service

The Problem

Most agent memory implementations stop at "store and retrieve." You save a fact, you search for it later. But as your memory corpus grows past a few hundred entries, raw retrieval isn't enough — you need the system to understand what it knows.

Over the past three releases (v10.55.0 → v10.57.3), mcp-memory-service crossed that threshold: from a storage layer to an intelligence tier.

What Shipped

Entity Extraction (v10.55.0, PR #868)

The memory service now automatically identifies entities (@mentions, #tags, project names) in stored memories and links them in a knowledge graph. This enables queries like "show me everything connected to @doobidoo" or "what decisions reference #authentication" — without manual tagging.

The design is deliberately conservative: regex-based NER (no LLM required), opt-in via MCP_ENTITY_EXTRACTION_ENABLED, integrated into the existing maintain cycle as Step 5.

Insight Cards (v10.55.0, PR #869)

After entity extraction runs, a second pass detects patterns, trends, and gaps across your memory corpus:

• Patterns: "You've stored 12 memories about authentication bugs in the last 30 days"
• Trends: "Your decision-making shifted from X to Y over time"
• Gaps: "You have 5 decisions about caching but no implementation follow-up"

These surface as memory_type='insight' entries — searchable like any other memory, but generated autonomously.

tag_match AND/OR Filtering (v10.55.0, PR #904)

A small but high-impact change: memory_search and memory_list now accept tag_match="all" for AND filtering. Previously, searching for tags=["project-x", "decision"] returned anything tagged with either — now you can require both. Essential for multi-project workflows.

Session Chunking (v10.56.0, PR #912)

memory_store_session now automatically chunks long conversations (>4000 chars) into semantically coherent pieces. Each chunk is independently searchable while maintaining session-level coherence via session:<id> tags.

Real-World Numbers

From a production deployment (~2,700 memories, sqlite-vec backend, 3 machines synced):

• Entity extraction: precise on @mentions/#tags, useful for cross-referencing projects across sessions
• Insight cards: maintain cycle completes in <10s, surfaces actionable gaps
• tag_match AND: reduced search noise significantly for scoped queries
• Harvest with locale patterns (fork): 75 candidates/session vs 9 English-only (8x improvement)

What's Next

• Phase 4: Temporal Contradiction Detection — identifying when newer memories contradict older ones (design discussion open)
• Multilingual harvest patterns — YAML-based locale plugins so non-English users get the same extraction quality (Design: Multilingual pattern plugins for harvest extractor #928)
• Kiro CLI support — extending harvest to parse additional MCP client formats (Design: Kiro CLI session format support for memory_harvest #927)

Try It

npx -y @doobidoo/mcp-memory-service

Or with entity extraction + insight cards enabled:

export MCP_ENTITY_EXTRACTION_ENABLED=true
export MCP_INSIGHT_CARDS_ENABLED=true

---

Contributors: @doobidoo (maintainer, entity extraction, insight cards), @filhocf (tag_match, session chunking, harvest locale plugins)

[doobidoo]

Thanks @filhocf — this is a solid draft and captures the arc well. Three corrections before I'd cross-post it anywhere:

1. Issue numbers are wrong
The draft references #927 (Kiro CLI) and #928 (multilingual harvest) — neither exists in the repo. The actual open issues are #908 (multilingual harvest) and #909 (LLM-based pattern discovery). Kiro CLI never got a tracking issue; the PRs (#892, #911) were closed without a follow-up. Either open a real issue for it or drop it from the roadmap section.

2. Attribution for entity extraction and insight cards
The footer reads: @doobidoo (maintainer, entity extraction, insight cards). That's not accurate — you wrote both #868 and #869. I squash-merged them directly and resolved the conflicts, but the implementation is yours. Please flip the attribution so you get the credit.

3. Locale plugins — let's revisit
I closed #913 a bit hastily. If you're willing to take another run at it, I'm open to merging a well-scoped locale plugin system — YAML-based config, sensible defaults, clear opt-in. Open a design issue so we can agree on scope before you write the code, and reference it from the blog's "What's Next" section so readers can follow along.

Fix those three and the post is good to go. Happy to cross-post to dev.to once they're addressed.

[filhocf]

Thanks for the review! Corrections applied:

1. Issue numbers fixed:

• Removed Design: Kiro CLI session format support for memory_harvest #927/Design: Multilingual pattern plugins for harvest extractor #928 (those are discussions, not issues)
• "What's Next" now references Support Kiro CLI JSONL format in memory_harvest #934 (Kiro CLI harvest — just opened), Multilingual harvest pattern support via locale plugins #908 (multilingual harvest), and LLM-based pattern discovery for harvest locale plugins #909 (LLM-based pattern discovery)

2. Attribution corrected:

• Footer now reads: @filhocf (entity extraction feat(reasoning): entity extraction and memory-entity linking (Phase 2, #732) #868, insight cards feat(consolidation): Insight Cards — automated pattern/trend/gap detection (Phase 3, #732) #869, tag_match feat: add tag_match parameter to memory_list #904, session chunking feat: automatic chunking for memory_store_session #912), @doobidoo (maintainer, squash-merge + conflict resolution, ongoing maintenance fixes v10.55.1→v10.57.3)

3. Locale plugins:

• Referenced the design issue process. "What's Next" links to Multilingual harvest pattern support via locale plugins #908 with note that a design issue is the prerequisite.

---

Updated "What's Next" section:

• Phase 4: Temporal Contradiction Detection — identifying when newer memories contradict older ones (design approved in Research & Design: Bridging Advanced Reasoning with MCP Memory Lifecycle #732, implementation next)
• Multilingual harvest patterns — YAML-based locale plugins so non-English users get the same extraction quality (tracking: Multilingual harvest pattern support via locale plugins #908, design issue pending)
• Kiro CLI support — extending harvest to parse additional MCP client formats (Support Kiro CLI JSONL format in memory_harvest #934)

Updated footer:

Contributors: @filhocf (entity extraction #868, insight cards #869, tag_match #904, session chunking #912), @doobidoo (maintainer, squash-merge + conflict resolution, ongoing maintenance v10.55.1→v10.57.3)

---

Ready for cross-post once you confirm these are good.

[doobidoo]

All three corrections land well — attribution and issue numbers are now accurate.

One small update before you post: the "What's Next" section is already behind. Since this morning:

• Multilingual harvest (Multilingual harvest pattern support via locale plugins #908) — shipped in feat(harvest): locale-based pattern plugins for multilingual extraction #935, merged today. Change "design issue pending" → "shipped in v10.57.x"
• Kiro CLI (Support Kiro CLI JSONL format in memory_harvest #934) — PR feat(harvest): Kiro CLI session format support in TranscriptParser #936 is open and under review. Change "tracking: Support Kiro CLI JSONL format in memory_harvest #934" → "PR feat(harvest): Kiro CLI session format support in TranscriptParser #936 in review"

Once those two lines are updated the post is ready. Go ahead and cross-post to dev.to.