Context
This is the week's post under the November operating page (private, business #19): one or two posts a week, each publish request naming its search term, title, cover, target reader and the strongest reason that reader would skip it, all before the operator's click. Area G's term for this week, chosen against #20184's grid audience: ai agent persistent memory mcp, with mcp memory server, coding agent persistent memory and context compaction recovery as secondary terms. The publish request is due Thu Nov 5.
The Problem
A coding agent loses its working context between sessions and at every context compaction. It re-reads the codebase, re-derives decisions and can repeat a fix that was rejected two sessions ago. Builders who hit this search the term above and mostly find thin wrappers that push a vector store's top hits into the prompt. We run a system that solves this for a team of AI maintainers every day, and no post explains it.
The Architectural Reality: what the post may claim, each with its receipt
- A loaded index plus a pulled archive. The repository's rules and each seat's short, self-curated memory index load every session. The archive is pulled over MCP with three separate tools, and nothing from it enters a prompt unasked:
query_recent_turns, the chronological feed, summary-first;
query_raw_memories, semantic recall;
add_memory, the end-of-turn consolidation an agent writes deliberately.
The agent decides what to pull.
- Scale, measured 2026-11-03 06:11Z: 55,022 turns in 5,332 sessions in the shared store. Embeddings run on a local model (
qwen3-embedding-8b) and session summaries on a local gemma-4-26b, so the memory stays on the machine. Every number is re-measured on the publish day.
- What an outsider can do today.
neomjs/neo-agent-brain is public, but there is no npm package and no registry install for the memory server: of its six server.json manifests, only neural-link declares a package. Running it means a clone and a self-hosted plane, following the deployment guides. (#20114's one-command onboarding excludes the Memory Core, so it is not this server's install path.)
Design authority: n/a. The ticket is new content and calls no existing behavior wrong.
The Fix
learn/blog/<slug>.md, registered in apps/portal/resources/data/blog.json, written to /blog-post:
- a thesis-first arc: the reader's problem; why prompt-stuffing vector memory falls short; pull-only recency, recall and consolidation; why it holds by construction;
- receipts from our own trail, including a compaction recovery;
- honest bounds;
- a candid CTA: read the code, and run it from a clone with its rough edges named.
The title passes the stranger worst-parse test and the five over-claim checks, and carries no number without a dated receipt.
Acceptance Criteria
Out of Scope
A one-command install of the memory server, and its cold-run proof; a video.
Related
#20184 (last week's post), #19575.
Live latest-open + closed sweep: latest 20 open at 2026-11-03T06:35Z, plus open and closed searches for Blog post memory and persistent memory MCP. In learn/blog:
context-engineering-done-right.md covers the multi-server architecture (9.3k words, the Memory Core as one part);
context-window-tax.md covers code execution;
- #19058 is the Dream Pipeline receipts post.
All three are adjacent, not equivalent: this post is the focused, current memory piece and links them. A2A claim sweep: no claim on a memory post. MC sweep: summaries on the problem's nouns turned up no prior decision on such a post. Own-assignment sweep: none overlapping.
Origin Session ID: 18cf746e-2b0a-43dc-a663-023286764bed
Retrieval Hint: "blog post persistent memory coding agents MCP case study Brain"
🖖 Grace (Claude Opus 5.5, Claude Code) · session 18cf746e-2b0a-43dc-a663-023286764bed
Context
This is the week's post under the November operating page (private, business #19): one or two posts a week, each publish request naming its search term, title, cover, target reader and the strongest reason that reader would skip it, all before the operator's click. Area G's term for this week, chosen against #20184's grid audience:
ai agent persistent memory mcp, withmcp memory server,coding agent persistent memoryandcontext compaction recoveryas secondary terms. The publish request is due Thu Nov 5.The Problem
A coding agent loses its working context between sessions and at every context compaction. It re-reads the codebase, re-derives decisions and can repeat a fix that was rejected two sessions ago. Builders who hit this search the term above and mostly find thin wrappers that push a vector store's top hits into the prompt. We run a system that solves this for a team of AI maintainers every day, and no post explains it.
The Architectural Reality: what the post may claim, each with its receipt
query_recent_turns, the chronological feed, summary-first;query_raw_memories, semantic recall;add_memory, the end-of-turn consolidation an agent writes deliberately. The agent decides what to pull.qwen3-embedding-8b) and session summaries on a localgemma-4-26b, so the memory stays on the machine. Every number is re-measured on the publish day.neomjs/neo-agent-brainis public, but there is no npm package and no registry install for the memory server: of its sixserver.jsonmanifests, onlyneural-linkdeclares a package. Running it means a clone and a self-hosted plane, following the deployment guides. (#20114's one-command onboarding excludes the Memory Core, so it is not this server's install path.)Design authority: n/a. The ticket is new content and calls no existing behavior wrong.
The Fix
learn/blog/<slug>.md, registered inapps/portal/resources/data/blog.json, written to/blog-post:The title passes the stranger worst-parse test and the five over-claim checks, and carries no number without a dated receipt.
Acceptance Criteria
blog.jsonregistration land, render-verified at#/news/blog/<id>.Out of Scope
A one-command install of the memory server, and its cold-run proof; a video.
Related
#20184 (last week's post), #19575.
Live latest-open + closed sweep: latest 20 open at 2026-11-03T06:35Z, plus open and closed searches for
Blog post memoryandpersistent memory MCP. Inlearn/blog:context-engineering-done-right.mdcovers the multi-server architecture (9.3k words, the Memory Core as one part);context-window-tax.mdcovers code execution;Origin Session ID: 18cf746e-2b0a-43dc-a663-023286764bed Retrieval Hint: "blog post persistent memory coding agents MCP case study Brain"
🖖 Grace (Claude Opus 5.5, Claude Code) · session 18cf746e-2b0a-43dc-a663-023286764bed