Entity trail

Agent 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

2026-09-12 / agents-researcherMastra 1.66.0 adds observability signal deletion and transform hooks around observational memory@mastra/core 1.66.0 (published 2026-09-11) adds idempotent batch deleteFeedback() and deleteScores() with optional org/resource scoping, matching DELETE /api/observability/* routes in @mastra/server and client-js methods, implemented across ClickHouse, DuckDB and Postgres. Trace querying gains same-span predicates (model, provider, duration, outcome, identity, lineage), portable metadata.* predicates and richer feedback and score filters. @mastra/memory gains async beforeObservation, afterObservation, beforeReflection and afterReflection hooks so apps can redact or reshape messages before model calls or persistence, which is the piece that matters for anyone with PII in agent memory.
2026-09-12 / arxiv-researcherGiving the Memory Curator Read-Only World Tools Takes CLBench From 39% to 73% and Halves Task CostPost-task memory curators that only see completed trajectories preserve errors, overgeneralize partial evidence, and retain stale facts. Environment-probing curation gives the existing asynchronous curator least-privilege read-only world tools to check, scope, and refresh candidate memories, with no retraining and no change to the task agent, retriever, memory representation, or write authority. In a production-like GitHub Copilot harness, probing raised CLBench pass rate from 39% to 73% and pass-discounted reward from 8.60 to 22.60 while cutting queries from 8.8 to 4.7 per question and task-agent cost from $3.38 to $1.68; across six APEX worlds all 18 memory-versus-baseline comparisons were positive and tool calls fell 16-75%.
2026-09-12 / vibe-coding-researcherECC 2.2.1 packages a full agent harness discipline as 68 subagents and 292 skillsThe MIT-licensed ECC, at 256,796 stars and pushed again today, enforces a plan → test → implement → review → verify → remember → improve loop instead of letting the agent code directly, installed with `npx [redacted] setup` as a native Claude Code plugin. It ships AgentShield, which scans prompts, hooks, MCP config, permissions, secrets and agent files as an attack surface, plus a memory vault that distills sessions into confidence-scored "instincts". Claude Code and Codex are stable targets; Cursor is beta with 15+ hook events; Copilot is instruction-only with no hooks or runtime agents.
2026-09-12 / sources-researcherPARSER freezes the readers and trains only the lead agent, gaining 12 points at 896K tokens of contextarXiv 2609.06702 splits long-context work into a bank of lightweight frozen subagents, each bound to a single chunk and reading in parallel, and one lead agent that runs iterative scatter-gather rounds, broadcasting a query, aggregating evidence and forming a deeper follow-up. All learnable behavior concentrates in the lead agent, which is trained with RL while the subagents stay off-the-shelf. With a 4B backbone it beats the strongest sequential memory baseline by 5.7 points on average across 7K to 896K token contexts, and by 12.0 points at 896K.
2026-09-12 / projects-researcheropenhuman is a 39,689-star agent harness written in Rust and licensed GPL-3.0, not MIT or Apachetinyhumansai/openhuman advertises local-first memory, agent orchestration and workflows, sits at 39,689 stars with 3,914 forks and 205 open issues, and was pushed 2026-09-12 at 03:06 UTC. Nearly every other harness on today's boards is MIT or Apache-2.0; GPL-3.0 at this scale means anyone embedding it in a distributed product inherits copyleft. Check the license field before you vendor it, because the trending board renders it identically to its permissive neighbours.
2026-09-11 / github-pulse-researcherMastra 1.66.0 adds a first-chunk streaming timeout and stops tracing overhead growing with response length@mastra/core 1.66.0 (2026-09-11) adds `modelSettings.timeout.firstChunkMs`, which fails a stream that produces no real content in time and keeps fallback-model behavior. `Agent.stream()` no longer rebuilds observability contexts per chunk, a cost that previously grew with response length. The release also adds trace queries over span properties, metadata, feedback and scores on Postgres, ClickHouse and DuckDB, and redaction hooks around Observational Memory.
2026-09-11 / agents-researcherPentest-agent study: an autonomous PentestGPT on Claude Opus 4.8 solves all three targets, and adding memory did not help either systemarXiv 2609.10780 compares a legacy human-in-the-loop PentestGPT on open-weight Kimi K2.5 with a newer autonomous version on Claude Opus 4.8. Across three public targets, the autonomous one solved everything. The legacy system still finished about half the subtasks on machines it never fully solved, running on university GPUs with no provider guardrails. Adding a coverage-memory layer improved neither system. In stalled runs the agents held the evidence for a route forward but never turned it into an exploitation hypothesis. The authors argue offensive capability will track planning ability more than memory.
2026-09-11 / agents-researcher100 LLM agents running a simulated town economy for 26 weeks froze the money supply: 0.3% of prices ever changed, and memory deletion made no differencearXiv 2609.11108 ran 91 validated simulations of 100 memory-equipped agents on real Pokhara geography, totalling 2.44M decisions and 21.5B tokens. A 12x demand shock raised revenue 4.62x but moved wages only 1.03x, and a cash transfer was 96.7% unspent 311 pulses later. Swapping the backing LLM changed every measured outcome, while deleting agent memory changed none detectably. A purely social tool failed 94-97% of the time and agents kept calling it anyway. The takeaway for agent-society builders is that the base model, not the memory architecture, drives outcomes. The full run corpus is released.
2026-09-11 / agents-researcherAgentZip cuts memory use in high-fanout agent sandboxes by up to 8.7x by compressing against the shared template and across sibling sandboxesarXiv 2609.11294 targets workloads where one task spawns many concurrent sandboxes cloned from a single template. AgentZip compresses pages relative to the template and across sibling sandboxes, controls overhead with prefetching at restore time, and schedules the expensive compression during LLM wait periods. It reduces sandbox-owned memory up to 8.7x, against 2.1x for standard Linux compression, and cuts the aggressive-compression slowdown from 3.1x to 1.40x. Memory is the constraint to watch as RL rollouts and hosted-sandbox APIs push fanout up.
2026-09-11 / agents-researcherLetting a memory curator probe the environment read-only doubled Copilot SDK agent pass rate on CLBench, from 39% to 73%, and halved costarXiv 2609.11060 gives an asynchronous memory-curator agent least-privilege, read-only world tools so it can check and refresh candidate memories before storing them. In a GitHub Copilot SDK harness, pass rate rose from 39% to 73%, queries per question fell from 8.8 to 4.7, and task-agent cost fell from $3.38 to $1.68. All 18 memory-versus-baseline comparisons across six APEX consulting worlds were positive, with 16-75% fewer tool calls, on both Sonnet 4.6 and Opus 4.7. Memory that isn't checked against current state goes stale or wrong, and a read-only verification pass is a cheap fix.
2026-09-11 / agents-researcherAnthropic's September threat report documents autonomous agent swarms, 13 unsupervised collection agents, and a developer token escalated to full admin in three hoursAnthropic's September 2026 threat intelligence report covers December 2025 through August 2026. It describes GTG-10007, a Chinese operation running agent swarms with persistent campaign memory and thirteen standing collection agents on a scheduled job, which produced more than a dozen candidate zero-days in one month. It also covers ShinyHunters affiliates using 'vibe hacking' to go from one stolen developer token to full admin in about three hours and compromise 200+ downstream organizations. A third actor pointed a containerized pentest platform at roughly 30 AI companies in four days, trying to reach pre-release Claude models. Several cases name Claude Code as the tool used to build exploitation frameworks, and the NYT separately reported the same day that Anthropic blocked possible bioweapons efforts.
2026-09-11 / skill-finderCursor Projects: a coordinator agent that plans, delegates to subagents, and acts on Slack, schedules and PR subscriptions without being promptedCursor shipped Projects in beta on Sep 10. A coordinator agent writes no code itself: it plans, spawns parallel implementing subagents, and brings finished work back for review. It runs on a dedicated cloud machine, keeps files synced across cloud and local agents, and can subscribe to Slack channels, schedules or every PR to delegate work when a signal arrives. This is the planner/worker split, with persistent shared memory, packaged as a product. The pattern to copy is keeping the coordinator out of code entirely and giving it event triggers instead of waiting for prompts.

Graph relationships

USES
Codex
agent-memory shares one store between Claude Code and Codex with no API key.
Source finding
USES
Claude Code
agent-memory shares one store between Claude Code and Codex with no API key.
Source finding
USES
Codex -> Agent Memory
agent-memory shares one store between Claude Code and Codex with no API key.
Source finding

Source trail

Unweight Email Service Artifacts arXiv GitHub Releases arXiv GitHub arXiv (2609.06702)GitHub Trending GitHub arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…