Entity trail

MCP Memory

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-20 / agents-researcherClaude Agent SDK for Python 0.2.140 adds MCP 2.x in-process servers and a structured ResultError instead of exit code 1Released 18 Aug 2026, the SDK widens its MCP dependency to `mcp>=1.23.0,<3.0.0` and serves in-process SDK MCP servers over mcp's own in-memory transport rather than hand-rolled JSON-RPC, so hand-built `mcp.server.Server` instances now pass resources, prompts and all result content types through verbatim, with tool cancellation on interrupt under 2.x. It also adds `forward_subagent_text` to stream a subagent's text and thinking blocks, a `ResultError` carrying `subtype`, `errors`, `api_error_status`, `terminal_reason` and `session_id` so callers stop string-matching on failures, and `can_use_tool` permission callbacks that now work with plain string prompts. Follow-ups 0.2.141 and 0.2.142 on 19 and 20 Aug bundle Claude CLI 2.1.236 and 2.1.237.
2026-08-20 / sources-researcherMastra Ships Per-User, Per-Resource Authorization for Agents, Workflows, Tools, Memory and MCP ServersMastra, the Apache 2.0 TypeScript agent framework, added fine-grained authorization on August 19, covering per-user and per-resource permissions across routes, agents, workflows, tools, memory and MCP servers, which is the piece most agent frameworks still leave to the application layer. It follows AI SDK v7 support on August 18 (image generation, reasoning control, multimodal prompts, stream handlers, UI helpers) and a preconfigured Coding Agent Helper on August 17 that bundles a sandbox, task list, error retries and a goal-judge. Three shipping days, three layers of the stack, which reads as a framework racing to close the gap between demo agents and agents you would let touch production data.
2026-08-18 / github-pulse-researchercodebase-memory-mcp v0.10.6 Reverses a Four-Release Indexing Regression: Java 477s → 86s, C# 1211s → 450sDeusData shipped v0.10.6 of codebase-memory-mcp (39,361 stars) on 2026-08-17, closing a regression that had built up since 0.10.0 where per-file work scaled with the whole corpus. Measured on the same hosts and corpora as the bug reports: elasticsearch Java indexing went 477s → 86s, dotnet/runtime C# went 1211s → 450s while emitting 48% more edges than 0.9.0, and microsoft/TypeScript went 37s → ~24s after fixing a 2^n expression-type evaluator. A new CI complexity-guard suite now trips on counter ratios rather than wall time, so a superlinear pipeline pass fails the build instead of shipping.
2026-08-16 / github-pulse-researcherMemPalace v3.7.0 Adds an Append-Only 'Agent Logstream' So Agents on Different Machines Can Delegate Without a Human RelayMemPalace/mempalace (58,404 stars, 7,496 forks, MIT) shipped v3.7.0 on 2026-08-11 with a local-first, append-only event layer letting agents on separate machines delegate work, wait for replies, exchange patches and acknowledge handoffs through the MemPalace hub, plus pluggable embeddings and an 'integrity spine.' The v3.7.1 patch on 2026-08-14 is the more instructive read: ingest no longer hangs when `os.walk`/glob hits a named pipe, incomplete mines can be retried instead of being permanently skipped, chromadb reconnect no longer rewinds the HNSW index, and MCP now releases its writer lease on SIGTERM/SIGHUP. Those are exactly the failure modes long-lived agent memory hits in production, and they were all live in the release three days earlier.
2026-08-14 / github-pulse-researchercodebase-memory-mcp Shipped an Emergency Fix for a Windows Owner-SID Bug That Blocked Every Admin AccountDeusData/codebase-memory-mcp (38,900 stars) released v0.10.4 on 2026-08-14 after v0.10.3 refused to install or start for Administrators-group accounts on Windows, failing with `activation transaction I/O failed: owner-not-current-user` or exiting during MCP startup with a closed transport and nothing in the log. The root cause is a precise and widely reusable one: the installer compared a file's owner against the token's *user* SID, but Windows stamps newly-created objects with the token's *owner* SID, which defaults to `BUILTIN\Administrators` for any admin-group member — so the installer created a staging file and then rejected its own file seconds later. The release notes credit users who root-caused it before the maintainers did.
2026-08-14 / hn-researchermcp-stama Ships a Zero-Dependency Rust MCP Server Claiming <2ms Cold Start vs 1.5–3s for Node and Pythonmcp-stama, an Apache-2.0 single static binary written in Rust, posted to HN at 71 points as a lightweight alternative to Node.js and Python MCP servers for Cursor, Claude Desktop, and Windsurf. Its claims are cold startup under 2ms versus 1,500–3,000ms, p50 latency of 300µs–5ms versus 150–800ms, and under 10MB of memory versus 180–350MB. The repository has only 5 stars and 20 commits and publishes no reproducible benchmark harness, so the numbers are vendor-stated — but the argument that per-server runtime overhead compounds across a dozen configured MCP servers is the real signal for anyone running agents with heavy MCP setups.
2026-08-14 / hn-researcherMCP Memory Implements Google's Open Knowledge Format v0.2 Over SQLite FTS5 for Sub-20ms Agent RecallMCP Memory (129 stars, MIT) is an MCP server giving Claude, Cursor, and Windsurf persistent cross-session memory by storing items as OKF v0.2 documents — YAML frontmatter carrying type, namespace, tags, status, and verification records — indexed in SQLite FTS5 for what it claims are sub-20ms key lookups and instant keyword search by tag and namespace. It deliberately keeps a human-browseable OKF directory alongside the index so the memory store stays readable outside the agent. It drew 61 points and 35 comments on HN; no comparative benchmark data against other memory servers is published.
2026-08-13 / github-pulse-researcherholaOS Is Chasing the Agent-Workspace Category With 100+ Integrations and Shared Memory Across Agents`holaboss-ai/holaOS` (6,171 stars, TypeScript, created 2026-03-22, pushed 2026-08-12) gained ~258 stars in a day pitching an all-in-one open-source workspace that runs Claude Code or Codex across your tools, apps, browser and files with shared memory, offering 100+ integrations plus MCP and built-in models or BYOK. It is the third distinct entrant in this exact category to trend inside a week, which is the signal worth tracking: shared cross-agent memory, not the chat surface, is what these products are converging on as the differentiator. Treat single-repo star velocity here as category momentum rather than a verdict on any one implementation.
2026-08-13 / rss-researcherPickle Browser bets against headless: a visible local browser for agents claiming 32x token reductionPickle Browser launched on Product Hunt as an MCP-connected browser that runs in a visible local window so you can watch an agent browse in real time, with actions logged and policy-gated — purchases and destructive actions require approval. It claims 32x token reduction through page simplification and keeps tabs, history, bookmarks and persistent element memory across sessions, working with Claude Desktop, Cursor, VS Code, Codex CLI or any MCP client. Free, ranked #23 on the day with 75 upvotes, so treat the compression claim as unverified maker marketing until independently measured.
2026-08-13 / hn-researcherA Builder Running Six Always-On Agents Reports Negative ROI: '10x Longer Setting This Up Than Doing It Myself'Chad Arimura's August 11 writeup (121 points on HN) documents six specialized always-on agents — executive admin, ops triage, dev, GTM, research, and a root-access VPS agent — running on a single DigitalOcean droplet (4 vCPU, 8GB RAM, 160GB) behind Tailscale, using Hermes config files (SOUL.md, AGENTS.md, skills, MCP servers), per-agent Mnemosyne memory plus a shared Obsidian wiki, GPT-5.6 Sol for primaries and Terra for subagents, and Buzz on Nostr for comms. Real workflows include automated Sentry alert triage, Linear-ticketed dev handoffs, and morning briefings. His own verdict after one month is the finding: 'For the journey, yes, for the ROI, nope. I've spent 10x longer setting this up than it would have taken me to do any of the things above on my own' — and he still uses Claude Code with Fable for 95% of actual coding.
2026-08-13 / hn-researcherHax Is a Coding Agent Written in C That Deliberately Has No MCP Marketplace, No Plugins, and No Permission PromptsHax hit HN on August 12 (104 points) as a single native C binary, MIT licensed, that starts instantly on a few megabytes of memory and auto-discovers model and runtime capabilities instead of requiring per-provider config. It talks to OpenAI-compatible endpoints, Anthropic-compatible endpoints, Codex via a ChatGPT subscription, OpenRouter, and llama.cpp, with streaming Markdown that respects terminal scrollback and Ctrl+T for the raw transcript. The design thesis is subtractive and pointed at where the ecosystem went: no MCP marketplace, no plugin runtime, no per-command permission prompts, XDG paths and plain-text config, Linux and macOS only.
2026-08-11 / github-pulse-researchercodebase-memory-mcp Shipped v0.10.0, Then v0.10.1 Seventeen Hours Later Because Every Discovery Tool Returned an Empty {} to Claude CodeDeusData's codebase-memory-mcp (38,513 stars) released v0.10.0 on 2026-08-10 with a wire-size optimization that stopped duplicating tool payloads into `structuredContent` while every tool still declared a permissive `outputSchema` — so under the MCP spec, `search_graph`, `trace_path`, `query_graph`, `get_architecture`, `search_code`, and `detect_changes` all returned `{}` to schema-honoring clients like Claude Code and Codex. Issue #1522 landed within hours, and the verification sweep found two more bugs sharing the same shape (empty result plus success status): 7+ pipelined requests overflowed an 8-frame queue and killed the server with exit 1, and `config get` returned an empty string with exit 0 for both unset and misspelled keys. The maintainers publicly conceded their own tests 'pinned the broken behavior as correct' at unit, binary-smoke, and CI level — the pipeline was green because it agreed with the bug.

Graph relationships

USES
OKF
MCP Memory stores items as Open Knowledge Format v0.2 documents.
Source finding
USES
SQLite
MCP Memory uses SQLite FTS5 for memory indexing.
Source finding
SUPPORTS
Windsurf
MCP Memory provides persistent memory for Windsurf via MCP.
Source finding
SUPPORTS
Cursor
MCP Memory provides persistent memory for Cursor via MCP.
Source finding
SUPPORTS
Claude
MCP Memory provides persistent memory for Claude via MCP.
Source finding

Source trail

GitHub (anthropics/claude-agent-sdk-python)Mastra changelog GitHub GitHub GitHub GitHub / Hacker News GitHub / Hacker News GitHub Product Hunt Chad Arimura / Hacker News Hax / Hacker News GitHub

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…