Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add U4CSolutions/MidMem --skill midmem-research-trackergit clone --depth 1 https://github.com/U4CSolutions/MidMemWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/u4csolutions/midmem/midmem-research-tracker)<a href="https://agentmods.dev/skills/u4csolutions/midmem/midmem-research-tracker"><img src="https://agentmods.dev/badge/skills/u4csolutions/midmem/midmem-research-tracker/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/u4csolutions/midmem/midmem-research-tracker"><img src="https://agentmods.dev/badge/skills/u4csolutions/midmem/midmem-research-tracker.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00160 | $0.01529 |
| Opus 5.5 | $0.00064 | $0.00612 |
| Sonnet 5.5 | $0.00032 | $0.00306 |
| Haiku 4.5 | $0.00016 | $0.00153 |
Grade A, and why
midmem-research-tracker scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 82 lines — stays where its author put it; the contents beside it link to each section on GitHub.
MidMem Research Tracker — ingestion → grounded ledger → roadmap-ready decisions
RESEARCH.md is the record of why the store is built the way it is: each paper paired with the
finding that matters and the decision it drove (or the reason it is parked). A weekly digest is an
LLM synthesis; the ledger is only trustworthy when every entry is checked against the source
document and the actual core code (DELEGATE-52 discipline — never grade the system from a summary).
Inputs
- The store: research-shaped sources in
state.db(typeresearch, weekly report / digest paths, arXiv-id file names). The deterministic intake list isnode scripts/research-sources.mjs(reads the ledger'sevaluated-throughmarker;--json,--all,--include-gone,--mark <ISO>). - The ledger:
RESEARCH.md(repo root). Marker line:<!-- research-tracker: evaluated-through=<ISO> -->. - Ground truth for "what exists":
packages/core/src/and the currentdocs/ROADMAP-*.mdstatus marks (✅ shipped / open increment numbers).
Procedure
- Intake.
node scripts/research-sources.mjs. For each unevaluated source read its grounding numbers first (summaryScore, quarantined counts come from the entry's own provenance).summaryScore≲ 0.4 or heavy quarantine → re-ingest permidmem-ingestbefore evaluating; a ledger built on a drifted extraction poisons the roadmap. - Extract, from the source document — not the store summary. For each paper the source
covers, capture:
title,arxiv_id,link(https://arxiv.org/abs/<id>; null if the source gives none — say so),finding(the empirical claim with its numbers), the source's own recommendation, and a relevance tag tied to ONE MidMem capability (retrieval · tiers/promotion · claims/contradictions · grounding · work-memory · concept graph · projection · bridge/ingest · bench/eval · governance/authority · project axis). Large reports → delegate extraction to a subagent that reads the file in full and returns JSON; never let it summarize from memory. - Ground-check each recommendation against the code. Targeted greps beat reading:
memFunction(typing) ·supersede|contradict|defer(claims) ·authority·communit|parent(graph) ·sufficien|progressive(gating) ·project(project axis) ·reembed|fallback(serving) ·bridgeSources|walkMarkdown(capture). Then place it against the roadmap: shipped (which ✅ increment), open (which #), or absent. - Feasibility rubric (deterministic, one line per paper).
- Fit — the subsystem it lands in (from step 2's tag).
- Effort — S (one module + smoke) / M (module + surfaces + docs) / L (schema or lane change).
- Discipline check — passes zero-deps, deterministic (no LLM in verifier/tagger/promotion paths), pure-core (no stack names), grounding-before-persist, maintain re-entrancy? Any "no" is a blocker to name.
- Evidence — numbers from a benchmark / a controlled result / prose only.
- Verdict — ADOPT NOW (serves an in-flight increment or the current build update and is S/M with a clean discipline check) · BACKLOG (real gap; gets or keeps a roadmap #) · VALIDATION (confirms a shipped design — cite it under that entry, it is roadmap evidence) · NOT ADOPTING (say why: contradicts a non-goal, filesystem-as-truth, LLM-judged promotion…).
- Write the ledger. Dated
## YYYY-MM-DD — <theme>sections, newest at top, one entry per ADOPT NOW / VALIDATION paper with Paper · Finding · Decision · Validation (validation = the smoke assertion, bench metric or deterministic check that proves the safeguard holds). BACKLOG papers go to the Backlog table (paper · finding · candidate · roadmap # · effort · blocker). NOT ADOPTING papers go to the Not adopting list with the reason. Every arXiv id links. Then advance the marker:node scripts/research-sources.mjs --mark <latest ingested_at evaluated>. - Roadmap consistency. A BACKLOG row without a roadmap # gets a new row in the current
docs/ROADMAP-*.md(same table format: increment · driver · principle served). An ADOPT NOW verdict does not open code — hand it tomidmem-devwith the ledger entry as the spec. - Record. Commit the ledger + roadmap delta (
docs: research ledger <date> — N papers, A adopt / B backlog / V validation), push, and store one lesson per genuinely new rule viamidmem-record. If a changelog exists, one line with counts and the marker.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago First seen · 82 lines · 160 tokens per session scan A 542aa925a556
midmem-research-tracker is a skill published in the GitHub repository U4CSolutions/MidMem (2 stars, last pushed 3d ago), licensed Apache-2.0. It adds 160 tokens to every session and 1,529 once invoked, about $0.0006 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-10.
Other skills, from other repositories
ori-memory
Persistent agent memory with learning retrieval. Knowledge graph on markdown files — capture insights, decisions, research, and learnings during work, then retrieve them weeks or months later. Use when knowledge is too valuable to lose but too much to inject into every prompt.
lemmalog
Externalize working memory and logical state into the lemmalog Datalog engine (MCP). Use for ANY multi-step task where state should outlive one context window or span agents: long investigations, debugging sessions, audits, multi-agent searches, systematic explorations, planning with many interdependent constraints…
braindb
Memory recall and save. Use at every conversation start and whenever the user shares personal information, expertise, project context, preferences, or decisions worth remembering long-term.
remnic-memory-workflow
Shared memory workflow for Claude Code agents connected to Remnic — recall before acting, observe during work, remember at the end. Trigger phrases include "what do you remember about", "save this for later", "any context from last time".
braindb-agent
Persistent memory across sessions via the BrainDB agent. Use at conversation start and whenever you need to recall what you know about the user or save new information to long-term memory.
remnic-recall
Search Remnic memories by natural-language query. Trigger phrases include "what do you remember about", "recall anything on", "have we discussed".