# Reported issues for neurodivergent-memory Pod holds 13 of 13 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user. Back to [neurodivergent-memory](/mcp/neurodivergent-memory). ## Most discussed ### Documentation audit: ensure MCP-first installation & clarify marketplace vs registry Intent: Ensure repository documentation is clear, accurate, and emphasizes MCP-first installation and configuration for the neurodivergent-memory server. This ticket scopes a full documentation audit across root docs and the `docs/` folder and will produce a short README patch and any further doc fixes needed. Scope: - Audit `README.md` and root-level docs (CHANGELOG.md, RELEASE_READY.txt, MARKETPLACE_SETUP_GUIDE.md, etc.). - Audit `docs/` folder including `marketplace-overview.md` and `ne [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/134) · 2026-05-19 · closed · 1 comment ### [chore] Modular core extraction — create src/core/ foundation (incl. epistemic_status) ## Summary Extract shared domain types, error taxonomy, and structured logging into a `src/core/` module directory. This is the prerequisite for all other v0.2.0 implementation work. ## Background Currently everything lives in `src/index.ts` (~700 LOC). Phase 1 of the architecture plan introduces `src/core/` without breaking the existing single-file integration point — `src/index.ts` remains the primary server entrypoint and imports from `src/core/`. ## Tasks - [ ] Create `src/core/` direct [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/14) · 2026-03-30 · closed · 1 comment ### [Rolodex Search & Creation] Task 5: /search route, pinned by a contract test Part of #162. Plan task 5. Read-only `GET /search?q=…` on the bridge, calling `search_memories` through the existing `runMcpTool` so the UI sees exactly what an agent would — same BM25, same tie-breaks — rather than a second divergent client-side filter. The contract test is the point: it runs the REAL tool against a seeded temp store and asserts the parser still recovers what it stored. A formatting change to the tool then fails CI instead of silently breaking search in production. If the pa [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/167) · 2026-08-02 · closed · 0 comments ### [Remote HTTP / Service Auth] Implement HTTP service mode Implement the core hosted HTTP service runtime for track B.`n`nDeliverables`n- service entrypoint and HTTP routing for MCP operations`n- explicit compatibility mode with existing stdio runtime`n- deterministic error contracts and observability for remote failures`n`nValidation`n- integration tests for service mode request handling [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/152) · 2026-07-08 · open · 0 comments ### [Remote HTTP / Service Auth] Design PAT/OAuth architecture Feature track B (separate from Google Drive track): hosted HTTP service access with PAT and/or OAuth auth. Goal - Deliver remote HTTP service mode with production-grade authentication (PAT and OAuth), independent from Google Drive-backed persistence decisions. Subtasks - [ ] #152 - [ ] #153 - [ ] #154 - [ ] #155 [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/151) · 2026-07-08 · open · 0 comments ### [Remote HTTP / Google Drive] Implement Google Drive storage/auth adapter Implement Google Drive-backed remote persistence for feature track A.`n`nDeliverables`n- adapter for reading/writing memory snapshot/WAL via Drive API`n- auth/token flow for Drive access`n- locking/synchronization strategy to prevent corruption/races`n`nValidation`n- conflict and recovery test scenarios`n- security review of token handling and permission scope [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/149) · 2026-07-08 · open · 0 comments ### [Remote HTTP / Google Drive] Implement HTTP transport mode Implement HTTP transport mode needed for remote access in the Google Drive feature track.`n`nDeliverables`n- HTTP server mode alongside stdio mode (no regression to stdio)`n- endpoint routing for MCP operations`n- request/response error contract aligned with current server behavior`n`nValidation`n- transport integration tests`n- local docs for launching in HTTP mode [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/148) · 2026-07-08 · open · 0 comments ### [enhancement] Add weak-client recovery shims: list_tools mirror and typo-tolerant no-result search assist ## Problem Yorkz/Cline dogfooding surfaced two recurring weak-client recovery gaps: 1. Some clients / weaker models fail to reason over native MCP tool discovery reliably and guess commands like `list_tools` 2. Single-token search near-misses can return a blank miss even when the intended project or token is obvious These gaps waste retries and make the server look less capable than it is. ## Evidence Supporting memories: - `memory_436` — grouped feature capture for cross-client ergonomics [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/92) · 2026-04-06 · closed · 0 comments ## Most recent ### [enhancement] Add recent-memory recovery hints to connect_memories NM_E004 errors ## Problem Yorkz dogfooding showed a recurring ergonomics failure around `connect_memories`: - a memory write succeeded - a follow-up `connect_memories` call failed on an exact-id miss - the agent had to do extra `list_memories` / `search_memories` round-trips to recover This is especially costly for weaker or rate-limited clients. ## Evidence Supporting memories: - `memory_405` — Yorkz-side proposal for recent-memory hints in the error payload - `memory_407` — project-level server proposal [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/91) · 2026-04-06 · closed · 0 comments ### Feature: Allow import_memories to load directly from snapshot file path ## Problem import_memories currently requires clients/agents to send a full entries array payload. In practice, many users already have a persisted snapshot (memories.json) on disk and want to bootstrap another server instance from that file. Today this forces clients to: 1. Read the entire file 2. Parse and transform records to import schema 3. Send a large token-heavy payload over MCP This is inefficient for agent workflows and unnecessary when the server can read local files directly. ## [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/49) · 2026-04-01 · closed · 0 comments ### v0.1.8 memory lost between sessions. As seen in the below PR text during my last experiment, memory was lost between conversations. ``` # AI-Orchestrated Development Experiment: Neurodivergent-Memory as Cognitive Substrate ## Overview This pull request was entirely conceived, planned, implemented, and submitted by an autonomous AI agent (Cline) using the **neurodivergent-memory MCP server** as its cognitive orchestration layer. The only human involvement was initiating the kanban-style workflow session. Every commit, every file [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/25) · 2026-03-31 · closed · 0 comments ### [feat] Structured logging and error taxonomy — pino logger, NM_E001–NM_E030 error codes ## Summary Replace all `console.log` / `console.error` calls with a structured JSON logger, and establish a stable error code taxonomy (`NM_E001`–`NM_E030`) that every failure path references. ## Background Current error handling: bare `process.stderr.write(...)` strings scattered through `src/index.ts`. No structured fields, no stable codes, no distinction between recoverable and unrecoverable failures. MCP tool error responses return raw strings. This blocks operator automation and monitori [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/17) · 2026-03-30 · closed · 0 comments ### [feat] Async write mutex and concurrency safety — queue, backpressure, WIP guardrail ## Summary Implement a true async write mutex that serializes all mutating MCP tool calls, add backpressure signaling under contention, and document the safe concurrency model. ## Background The current `saveChain` promise serializes disk writes but **not** in-memory mutations. Concurrent `CallToolRequest` events (store + delete arriving simultaneously) can interleave in-memory operations and produce inconsistent state before any save occurs. ## Tasks ### Async Write Mutex - [ ] Implement [Read the thread](https://github.com/jmeyer1980/neurodivergent-memory/issues/16) · 2026-03-30 · closed · 0 comments The remaining reports are on [the project's issue tracker](https://github.com/jmeyer1980/neurodivergent-memory/issues).