AlphaSignal - agentmemory hits 11k stars, claims 92% fewer tokens for Claude/Cursor (May 18 2026)

Why this is in the vault

The top item - agentmemory, an Apache-2.0 open-source background memory layer for coding agents - is directly load-bearing for Ray's own context-management story. It claims 92% fewer tokens per session vs loading full context, 95.2% retrieval accuracy via hybrid (keyword + vector + knowledge-graph) search, SQLite-only (no external DB), and MCP wiring for Claude Code, Codex, Cursor, Gemini CLI, Cline, Windsurf. That is exactly the gap CLAUDE.md hard rule #4 codifies for this agent (don't burn parent context on large artifacts) and exactly what working-context.md does today as a hand-rolled durable scratchpad. If the claims hold, this is a candidate substrate for the next iteration of Ray's memory architecture - or at minimum a strong reference for what the open-source community converged on.

The Lior framing in the intro is also notable for the vault thesis-tracking: "The real bottleneck in agentic coding isn't model intelligence anymore. It's memory... We're entering the context engineering era." That maps cleanly onto the harness-thesis evidence already in the vault (Cobus Greyling, Harrison Chase, Akshay Pachaar, Addy Osmani, Zhou et al.) - one more validation that the field's center of gravity has shifted from weights to runtime.

Secondary signal worth tracking: the Tom Dorr "106 Claude Code skills across 15 professions" repo - direct adjacency to RDCO's own skill library and the Garry Tan "Thin Harness, Fat Skills" pattern. May be worth a sniff-test for skill-shape ideas (PR Description Writer, Debugging Log Analyser, Runbook Writer, Changelog Generator are the cited engineering bundle).

Mapping against Ray Data Co

CLAUDE.md hard rule #4 (route long artifacts through subagents): agentmemory is the inverse-side of the same problem - rule #4 prevents context bloat in-session; agentmemory prevents it across-session. Both attack "context rot." If we ever adopt agentmemory, the rule would still apply within a session, but the "every session resets" pain Lior names would be solved at the substrate level.
~/.claude/state/working-context.md (durable scratchpad): today's hand-rolled equivalent. agentmemory claims SQLite-backed structured storage + compression + retrieval that beats raw file recall. Worth a paper-trade: stand up agentmemory in a sandbox session and measure recall quality against working-context.md on the same task surface.
Thariq implementation-notes pattern ([[2026-04-15-thariq-claude-code-session-management-1m-context]]): Thariq's frame is "implementation notes externalize state so the next session doesn't have to reconstruct it." agentmemory automates that externalization. The question is whether automated compression preserves the right things - implementation notes are written by the engineer who knows what mattered; agentmemory compresses heuristically.
Generative-UI / founder-COO interface (synthesis-queue #5): if Ray's memory layer becomes a queryable service rather than a flat file, the founder-facing surface can render it (e.g., "show me everything Ray remembers about the Squarely launch"). Substrate matters for the interface.
Harness thesis arc: another data point that the harness-engineering community is iterating in public, and that the memory slot is one of the more contested boxes in the harness diagram. Connects to [[2026-04-12-cobus-greyling-harness-era-language-shift]], [[2026-04-12-harrison-chase-harness-blog]], [[paper-arxiv-2604-08224-agent-harness-study-2026-04-12]], [[2026-05-10-addy-osmani-agent-harness-engineering]].
State-as-moat ([[2026-04-13-jaya-gupta-ai-lock-in-state-moat]] + [[../08-tooling/rdco-state-ownership-architecture]]): open-source SQLite-only is the opposite of vendor lock-in. Adopting agentmemory keeps the moat on RDCO's side rather than leasing it to a memory-as-a-service vendor. That fits the state-ownership posture.

Verdict: medium-to-strong mapping. Top item is worth a hands-on paper-trade, not just a "noted." Skill-bundle item is a sniff-test. Other curation items skip.

Issue contents

Top item: Tom Dorr - 106 Claude Code skills, 15 professions (324 stars, MIT)

Install via /plugin marketplace add mohitagw15856/pm-claude-skills or per-bundle.
Engineering bundle: Debugging Log Analyser, PR Description Writer, Runbook Writer, Changelog Generator.
Claims 8-10 hr/wk saved.
RDCO action: scan engineering bundle for skill-shape ideas. Don't install wholesale (security-review default applies). If a specific skill looks compelling, copy the shape into a vetted version under ~/.claude/skills/. Logging as a Notion task.

Signals (one-liners)

HKUDS opens tool that turns any software into AI-agent-ready CLI (36k stars). Adjacent to harness/tool-dispatch thesis. Skip-but-note.
Buildkite paid placement - CI orchestration. Skip.
Dograh - open-source self-hostable drag-and-drop voice agent builder (1.8k stars). Adjacent to ElevenLabs/voice work but not load-bearing today. Skip.
Qwen3.6-35B MoE with quantized GGUF + speculative decoding (5.3k downloads). Local-model inference improvement. Skip-but-note - relevant if RDCO ever moves any workload local.

⚠️ Sponsorship

FriendliAI - paid inference-optimization guide. Standard sponsor placement; framing is "inference cost optimization via TTFT/TPOT measurement + quantization + speculative decoding tuning." Useful primer if RDCO ever runs inference at scale, but not load-bearing today.
Kepler - paid placement; "deterministic infrastructure for AI" framing, hiring NYC founding engineers. The Kepler intro tease is interesting: "Past 200k tokens, sourcing discipline erodes. Bigger context windows don't fix sourced reasoning. They break it sooner." That's the same context-rot point Thariq made and that motivates rule #4. Sponsor commercial intent is hiring + infra-positioning, no immediate RDCO action.
Buildkite - paid CI orchestration placement. Skip.

All three are standard AlphaSignal paid placements, disclosed. No ideological bias.

[[2026-04-12-cobus-greyling-harness-era-language-shift]] - the "context engineering era" framing Lior uses maps to this
[[2026-04-12-harrison-chase-harness-blog]] - "Your Harness, Your Memory" is the same thesis
[[paper-arxiv-2604-08224-agent-harness-study-2026-04-12]] - academic version of the harness/memory thesis
[[2026-04-10-akshay-pachaar-agent-harness-anatomy]] - memory slot in the harness diagram
[[2026-05-10-addy-osmani-agent-harness-engineering]] - most recent harness-engineering reference; memory is one of the named slots
[[2026-04-13-jaya-gupta-ai-lock-in-state-moat]] - state-as-moat, why open-source SQLite-only matters
[[../08-tooling/rdco-state-ownership-architecture]] - RDCO's current state architecture; agentmemory is a candidate substrate
[[2026-04-11-garry-tan-thin-harness-fat-skills]] - Tom Dorr skill bundle is a literal instance of this pattern
[[2026-04-20-indydevdan-agent-experts-self-improving]] - the YAML expertise-file pattern is hand-rolled agentmemory
[[2026-01-27-every-compounding-engineering-intro]] - knowledge-compounding thesis; persistent memory is the substrate

alphasignal agentmemory 92 percent fewer tokens