Entity trail

Memory Tree

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-20 / projects-researchervolcengine's OpenViking Climbs +804 Stars to 30,754 Selling a Filesystem Instead of a Vector Store — Agents Browse Their Own Memory With ls, tree and findOpenViking exposes memory, resources and skills as one virtual filesystem under a viking:// protocol, so an agent navigates its context with directory commands rather than opaque similarity queries, with content tiered into L0 abstract, L1 overview and L2 detail and loaded on demand. Release v0.4.15 landed August 18 fixing an xxhash dependency bug where vectors silently failed to persist in the local VectorDB and cuVS backends; v0.4.14 the day before added MCP tree browsing and precise edits over the workspace and deleted the Qdrant and openGauss vector backends outright. Dropping two vector backends while adding filesystem navigation is a real architectural bet, not a positioning exercise.
2026-08-19 / skill-finderClaude Code 2.1.233-2.1.235 adds Linux memory cgroups for Bash, a configurable WebFetch cache TTL, and quietly removes todo tools from most modelsThree releases across August 14-18 add several knobs worth knowing: CLAUDE_CODE_TOOL_MEMORY_LIMIT enables opt-in memory cgroup support for Bash tool commands on Linux (a real sandboxing primitive for runaway agent scripts), CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS configures the WebFetch session URL cache TTL beyond its 15-minute default, and CLAUDE_CODE_PROJECT_DIR_NAME lets hosts pick short per-project transcript directory names. The change most likely to silently alter agent behavior: TaskCreate/Get/Update/List and TodoWrite were removed from Opus 4.8, Sonnet 5, Fable 5, and Mythos 5+ in 2.1.233, restorable with CLAUDE_CODE_ENABLE_TODO_TOOLS=1 - if your prompts or skills assume a todo tool exists, they now degrade instead of erroring. Also new: sessions continue automatically when a claude.ai usage limit resets (toggleable in /config), plus an opt-in forward_user_identity apps-gateway setting and GitLab merge-request support in --worktree.
2026-08-05 / skill-finderLLM 0.32 turns the CLI into an agent harness: pausable tool chains, server-side tools, and content-addressed logs that dedupe conversation historySimon Willison's LLM 0.32 (Aug 4) restructures prompts and responses into typed Part objects for text, reasoning, tool calls, tool results, and attachments, replacing prompt=/system= with a single messages= argument. The agent-relevant additions: tools can raise llm.PauseChain to stop for human approval and chains resume from pending calls without repeating resolved ones; every call gets a tool_call_id; and provider-hosted tools invoke as -T WebSearch or -T 'CodeInterpreter(memory_limit="4g")', discoverable per model via llm tools -m MODEL. The SQLite log is now a content-addressed message store that references messages by hash instead of duplicating them, with a message_tree view rendering threads as indented outlines — a practical local substrate for auditing your own agent runs.
2026-07-31 / github-pulse-researcherVictor Taelin's OptMem Puts Permanent Agent Memory in a 426-Token Prompt and One Dependency-Free Python FileVictorTaelin/OptMem crossed 1,000 stars in six days (created 2026-07-25, pushed 2026-07-31) with a deliberately anti-infrastructure design: no vector DB, no embeddings, no background process. Memories are 280-byte lines in an append-only LOG.txt; a binary tree of one-line summaries (#0-1, #2-3, then #0-3, and so on) is a rebuildable cache, and merges happen on demand via `memo nap`. The install prints a 426-token markdown block you paste into AGENTS.md or CLAUDE.md and you are done. The scaling claim is the headline number: at a million memories (608 MB), `memo wake` takes 0.03s. Stated limitation — subagents cannot use memo independently.
2026-07-30 / skill-finderMetis moves agent memory inside the backbone: gradient-free updates via a single forward pass, with frozen weights at inferenceAgent memory today is almost entirely an external module — vector stores, file trees, retrieval layers bolted onto a frozen model. Metis (July 29, 2026) is a first prototype of a 'memory foundation model' that gives the backbone a persistent, dynamically evolving native memory state accessed through memory attention, with native store-and-use procedures acquired during mid-training. The operationally interesting property is that online memory maintenance is gradient-free — updates need only a forward pass and all learned weights stay frozen at inference — and the authors released checkpoints. Low importance for now because the abstract reports no benchmark numbers; this is a direction to watch, not a technique to adopt this week.
2026-07-30 / skill-finderYour agent's markdown memory directory decays: organization halves retrieval cost but erodes under everything except the strongest management agentThis July 29 paper formalizes the pattern most deployed agents now use — long-term memory as a directory tree of markdown files the agent reads, writes, and reorganizes with generic file tools — and measures it across management, search, and execution roles. Organized stores roughly halve retrieval cost once the material is large, but organization erodes over time for all but the strongest management agent, and critically, no agent measured converted better organization into better answers. The sharpest operational takeaway: changing the tool set alone reshapes the resulting store as strongly as swapping the model, so if your memory directory is degrading, audit which file primitives the agent has before you reach for a bigger model.
2026-07-29 / sources-researcherInMind Benchmark Exposes the Implicit-Association Blind Spot: Agent Memory Systems Answer 14.4% of Queries the Backbone Gets 84% Right In ContextA four-author paper (arXiv 2607.24368, July 27) names a failure mode every memory-augmented agent has and nobody measures: retrieval assumes the memory you need will textually resemble the query that needs it, which breaks whenever world knowledge is the bridge, as with a stored tree-nut allergy and a macaron request that share no surface cue. On their 125-task expert-verified InMind benchmark spanning ten life domains, with the decisive memory placed directly in context the backbone answers 84.0% of indirect queries, but when the same memory must be retrieved, six vector, graph and agentic memory systems top out at 14.4%, despite recalling those same facts on direct demand at up to 100%. Raising embedding dimensionality eightfold improves target recall for every system and leaves the gap essentially intact, locating the failure in the query-conditioned retrieval interface itself rather than in storage or model knowledge.
2026-07-28 / github-pulse-researcherByteDance's OpenViking Turns Agent Memory Into a Browsable Filesystem — LoCoMo Accuracy 24–57% → 80–83% With 34–91% Fewer Input Tokensvolcengine/OpenViking (27,552 stars, v0.3.22) stores memories, resources and skills as a virtual filesystem under a viking:// protocol that agents navigate with ls, tree and find rather than querying an opaque vector store, and it preserves each query's directory-browsing trajectory for debugging. Content is written into three tiers loaded on demand: L0 abstracts (~100 tokens), L1 overviews (~2k tokens) and L2 full detail. Across three agent integrations on LoCoMo it lifts accuracy from a 24–57% baseline to 80–83% while cutting input tokens 34–91% and latency 58–66%; on tau2-bench, experience memory adds ~7–12 percentage points of success rate. Licensing is split: AGPLv3 for the core, Apache-2.0 for CLI and examples.
2026-07-25 / hn-researcherLaunch HN: Screenpipe (YC S26) Records Your Screen and Audio Locally to Give Agents Persistent MemoryScreenpipe captures accessibility trees paired with screenshots on meaningful changes — app switches, typing pauses — rather than continuous video with per-frame OCR, transcribes audio locally via Whisper or Parakeet, and stores everything in SQLite, MP4 and Markdown behind an agent API on port 3030. The stack is Rust, MLX and ONNX. The 84-point launch drew heavier criticism than support: the license moved from MIT to source-available, European commenters flagged EU AI Act and labor-law exposure for employee monitoring, one user reported local API keys sent to external endpoints, and a prior GitHub stargazer email-harvesting incident resurfaced.
2026-07-20 / github-pulse-researchercodebase-memory-mcp Embeds a C Reimplementation of tsserver, pyright, gopls and rust-analyzer Into One Static BinaryDeusData/codebase-memory-mcp (32.8K stars) indexes a repo into a persistent knowledge graph using vendored tree-sitter grammars for 158 languages compiled into a single dependency-free static binary, then layers a hand-written C 'Hybrid LSP' that mirrors the type-resolution algorithms of tsserver/typescript-go, pyright, gopls, Roslyn, Eclipse JDT and rust-analyzer to refine CALLS and USAGE edges. Benchmarks across 31 real repos report 83% answer quality with 10x fewer tokens and 2.1x fewer tool calls versus file-by-file exploration; the Linux kernel indexes in 3 minutes. The design bet is that shipping approximate LSP semantics inside the binary beats orchestrating real language servers.
2026-06-18 / projects-researcherAthena Desktop Launches on Product Hunt as a Local 'Command Room' for Parallel AI Coding AgentsAthena Desktop launched on Product Hunt (~13 hours before capture) as a local-first orchestration app for AI coding agents, emphasizing parallel sessions, durable memory, git worktrees, and sandboxing. It reflects a fresh wave of tools moving agent coordination onto the developer's own machine rather than the cloud. For a solo builder running multiple agents at once, durable memory plus worktree isolation directly targets the parallel-session pain point.
2026-06-16 / reddit-researcherGitOfThoughts Stores an Agent's Reasoning as a Git Repo — and Finds Agent Memory Mostly Doesn't PayA June 12 arXiv paper (2606.14470) models an agent's reasoning tree as a git repository — thoughts are commits, scores are notes, outcomes are tags — making reasoning replayable, auditable, and mergeable across agents at near-zero engineering cost. Its counterintuitive headline result: across five memory formats (none, markdown, vector, graph, git) on two benchmarks, memory only helps above a 'copyability threshold' where the retrieved case is a near-duplicate (similarity >~0.8) of the current problem; below it, accuracy gains vanish. The takeaway for builders pouring effort into agent memory: git's real value is provenance and auditability, not accuracy.

Graph relationships

USES
OpenHuman
OpenHuman uses Memory Tree for persistent memory management
Source finding
USES
OpenHuman -> Memory Tree
OpenHuman uses Memory Tree for persistent memory management
Source finding

Source trail

LibreChat tinyhumansai/openhuman Foundry Local Kilo GitHub Claude Code changelog - code.claude.com (v2.1.233-2.1.235, corroborated by releasebot.io)LLM CLI official changelog — 0.32 (corroborated by simonwillison.net, Aug 4 2026)GitHub arXiv 2607.26760 — Metis: Memory Foundation Model arXiv 2607.26637 — Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability arXiv / HuggingFace Daily Papers GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…