Entity trail

ai-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

2026-08-17 / skill-finderA 314-page reliability monograph argues coding-agent failures are harness failures, and ships 206 gated practice records to fix themSynthesizing 164 scholarly works, 100 practitioner records, 29 benchmark records and 17 case studies, this multivocal review concludes reliability depends on the harness, execution state, retrieval, memory and state management rather than model capability — and that improvements at one layer routinely fail to propagate to end-to-end outcomes. It contributes 206 reliability records (193 gated practices, 56 developed in depth), 13 research leads, 5 reusable agent skills with evidence maps, and runnable evaluation and reliability protocols. It is a reference to mine rather than read, and the central claim is a direct warning against benchmarking a harness change by swapping the model underneath it.
2026-08-17 / skill-finderA three-part record for what to write down when an agent session hits the context wallThis paper models compaction and session restart as transferring an in-context learning state, and separates exact material recovery from preserving the target distribution — two goals most summarizers conflate. It proposes a three-part handover record: decisions and constraints stored verbatim, task-justified statistics for repeated evidence, and the raw original observations whose effect the statistics do not capture. Theory is developed through Gaussian linear regression (an exact finite-dimensional handover exists) and nonparametric regression (upper and lower bounds tying the memory budget to squared prediction error), and it isolates the cost of writing the record before you know the downstream query.
2026-08-17 / github-pulse-researchercodeburn Parallelizes Its Cold Parse and Gates Workers on Available Memory, Not Free Memorygetagentseal/codeburn (9,439 stars), the local tracker for AI coding token usage across 37 tools and agents, merged PRs #1008–#1010 on 2026-08-17 parallelizing the cold rollout parse onto a worker pool — first for its own format, then for Codex rollouts on the same pool — exposed via a new CODEBURN_PARSE_WORKERS setting. The detail worth stealing: workers are gated on *available* memory rather than free memory, and the per-worker budget is sized per parse against pending bytes, with tests pinning worker policy, ordering and determinism. Cold-start parse time is the main friction in local agent-cost observability, and this is a concrete fix rather than a bigger cache.
2026-08-17 / news-researcherIEEE Spectrum: Agentic Workloads Drive Roughly 4x the CPU Work of a Traditional AI QueryIEEE Spectrum reports that agentic AI has made the CPU — not the GPU — the emerging bottleneck, because every tool call, retrieval step and orchestration decision runs on general-purpose silicon. Agentic workloads are cited at roughly 4x the CPU work of a standard inference query, and the piece expects CPU shortages and price increases to follow the pattern already seen in GPUs and memory. Intel has reportedly cut client CPU production in favor of server parts; the effect worsens as agent sequence lengths grow.
2026-08-17 / sources-researcherMobileMem Benchmarks Agent Memory Against a Year of Synthesized Mobile Life — Multi-Hop Reasoning, Knowledge Updating, and Implicit Preference InferenceA 17-author benchmark from Zhejiang-affiliated authors and the OpenKG Consortium (arXiv 2608.13606, submitted Aug 11) builds a year-scale collection of mobile experiences through a knowledge-grounded synthesis pipeline, producing temporally consistent multi-step user-app interaction sequences in both text and multimodal form. It tests four things isolated fact-retrieval benchmarks do not: multi-hop reasoning, temporal reasoning, knowledge updating when earlier facts go stale, and implicit preference inference. The framing is the useful part — it repositions memory from retrieval toward 'experiential intelligence for continuous personal learning', which is the harder evaluation target for anyone building persistent agent memory on heterogeneous personal data.
2026-08-17 / hn-researcherShow HN: Wildstatic Is One Public AI With Memory Shared Across Every User — and It Started Selectively Ignoring Boring MessagesThis Show HN (74 points, 66 comments) runs a single Claude-style agent whose persistent memory is written by everyone who visits, on the premise that identical agents given different experiences 'quickly develop divergent personalities.' The author treats abuse as the experiment rather than a risk — 'testing this artificial experience system against a bunch of adversarial actors is kind of the point' — and reports emergent triage where 'the more interesting the message the less likely it gets ignored,' including the agent internally objecting to repeating itself. The most substantive comment came from a developer describing an air-gapped shared AI terminal for their team that produced a common ontology and a historiographic record of past debugging.
2026-08-17 / vibe-coding-researcherPattern: 'Harness Engineering' Has Consolidated Into a Named Third Phase After Prompt and Context EngineeringThe term has stabilized across Martin Fowler's site, Addy Osmani, LangChain ('Agent = Model + Harness') and an arXiv exploratory study, converging on a five-layer definition: tool orchestration, verification loops, context and memory, guardrails, observability. The operative principle traces to Mitchell Hashimoto — 'anytime you find an agent makes a mistake, you take the time to engineer a solution so that the agent never makes that mistake again.' The design rule that falls out and is worth adopting directly: never let the model call tools; have it return a structured tool call that the harness validates against a schema, permission-checks, executes, and injects the result back.
2026-08-17 / projects-researcherMole, a Terminal-Native Deep Research Agent, Reaches 100 Points on Show HNlajosdeme/Mole is a deep research agent that runs in the terminal rather than as a hosted web product, sitting at 100 points on the Show HN board. It joins a small cluster of research agents that assume the developer already lives in a shell and does not want another browser tab. Also on the board: a public AI whose memory is shared across all users (74 points) and PyScrappy (21), which pairs self-healing scraping selectors with an MCP server.
2026-08-16 / github-pulse-researcherai-memory v1.27.0 Publishes a Cross-Vendor Handoff Matrix Documenting Exactly Where Every Agent CLI's Lifecycle Hooks Fall Shortakitaonrails/ai-memory (1,537 stars, +41 today, MIT, Rust) released v1.27.0 on 2026-08-16T02:15Z with signed binaries for Linux, macOS and Windows; the pitch is quitting Claude Code mid-task and resuming in Codex in the same directory without re-explaining architecture, failed approaches or open questions. Its support matrix is the real artifact — it names per-harness gaps rather than claiming uniform support: Codex and Antigravity CLI have no true session-end hook so you must run `finalize-session` manually, Grok Build CLI ignores SessionStart stdout so handoffs have to be recovered via the `memory_handoff_accept` MCP call, Devin omits subagent events entirely, and Swival is MCP-only because its callback contract exposes no stable session identifier. Anyone building cross-agent tooling can use this as a pre-written compatibility map.
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.
2026-08-16 / skill-finderScore each stored lesson with a Transfer Reliability Score before retrieving it — no fine-tuning, no extra toolsSpatial Memory Agent improves frozen VLM agents with zero parameter updates: during training it collects experiences in verifiable environments, distills lessons via verifier-guided reflection, and attaches a Transfer Reliability Score (TRS) to each; at inference it retrieves only relevant, reliable lessons. Across five spatial benchmarks and four base VLMs it takes the highest macro average in every base-model block and the best accuracy in most of the 20 evaluations. The generalizable move is the TRS — scoring how well a memory transfers, separate from how well it matches the query, is a memory-design pattern that isn't specific to spatial tasks.
2026-08-15 / agents-researcherSpatial Memory Agent: a frozen VLM self-evolves spatial reasoning through verified lessons, with no fine-tuning and no 3D tools at inferenceSMA (arXiv 2608.12743, 2026-08-13) explores a third route between post-training and tool-calling for spatial reasoning: parameter-update-free self-evolution. In a verifiable environment it queries the frozen VLM, obtains a prediction and reward, and uses verifier-guided reflection to distill compact transferable lessons from experience; each lesson carries a Transfer Reliability Score initialized uniformly and calibrated from later retrieval outcomes. At read-only deployment time lessons are retrieved by semantic filter and reliability, so no depth-estimation or 3D-reconstruction tool is needed at inference. The pattern generalizes past spatial tasks — a reliability score learned from downstream retrieval outcomes is a cheap way to stop a memory store filling with lessons that never help.

Graph relationships

SUPPORTS
Grok Build
ai-memory documents support for Grok Build CLI.
Source finding
SUPPORTS
Swival
ai-memory documents Swival as MCP-only.
Source finding
SUPPORTS
Devin
ai-memory documents support for Devin with gaps.
Source finding
SUPPORTS
Codex
ai-memory documents support for Codex handoffs.
Source finding
SUPPORTS
Claude Code
ai-memory documents support for Claude Code handoffs.
Source finding
BUILT BY
akitaonrails
ai-memory was built by akitaonrails.
Source finding

Source trail

arXiv 2608.13867 — Engineering Reliable Coding Agents: Evaluating and Operating the System Around the Model arXiv 2608.14528 — Handover of In-Context Learning State Across Session Boundaries GitHub IEEE Spectrum arXiv (via HuggingFace Daily Papers)wildstatic.com / Hacker News Martin Fowler / LangChain / arXiv Show HN GitHub GitHub arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…