ZinkLu had Claude review 1,050 of his agents' memory recalls one by one; about one in ten was confirmed used, and he drew practical rules for using memory better
ZinkLu (@ZinkLu2) uses Codex, Claude Code, OpenCode, Kimi Code and Hermes across a Mac and two Linux machines, all sharing one self-hosted Nowledge Mem memory store. After installing it he wanted to know how much it actually helped. He gathered 3,230 sessions, kept the 1,078 after memory was connected, defined a "recall window" as the stretch from one recall to the next (1,050 in total), and had Claude Opus 5.5 check each one against the prior context, the returned memory and what the agent did next; every "used" verdict needed a verbatim quote, checked by a script. Results: 101 windows confirmed used, 9.6% strict or 12.4% lenient; memory took a median 4.1% of context; empty results and errors were 28% of recalls. The details are the useful part: the first recall in a session was used 19.6% of the time versus 9.5% after; queries over 60 characters were used 18.2% versus 3.3% for under 15; retrying right after an empty result or error rarely helped; memories under a day old were used more than those over 30 days; and on the same 120 real queries, memory search gave directly usable results 25.8% of the time versus 6.7% for searching past sessions. We did not reproduce this analysis.
Visit site Nowledge Mem + Codex / Claude Code / OpenCode ↗- Pricing
- Freemium
- Region
- Global
- Where it's from
- Chinese community
- Added
- Nov 13, 2026
- Version we used
- Article 2026-11-12; Nowledge Mem self-hosted, sessions 2026-06-29 to 2026-10 (Claude Code 09-03 to 10-04), reviewed by Claude Opus 5.5
- Categories
- CodingResearchProductivity
- Original author
- ZinkLu @ZinkLu2 ↗
- Our test
- Not tested by us
- Cost / time
- Nowledge Mem's free plan stores up to 50 memories and works fully with your own model keys; Plus starts at $9.99/month. Reviewing a thousand windows uses a fair amount of model quota; the article gives no figure.
The task
You gave several coding agents shared memory and the recalls look busy, but you can't tell whether they prevent mistakes or just eat context.
Steps
- Collect agent session logs from every machine (Claude Code keeps only 30 days by default, so back them up early). From the date you started using memory, keep sessions with your prompts or memory calls.
- Find every memory call in the logs, remove duplicates from forked sessions, and cut a recall window from one recall to the next recall, context compaction or end of log.
- Have an AI reviewer judge each window: what was already in context, what memory returned, and whether later replies or actions used it. Require a verbatim quote for every "used" verdict and verify it with a script; give a sample to a second reviewer to check agreement.
- Compute adoption by agent, task type, recall order, query length, memory age and source, plus empty results, errors and context share.
- Re-run a batch of real past queries through each retrieval method (memory search, past-session search) and have a reviewer who doesn't know the tool names judge whether results are directly usable. Exclude memories written or edited after the original query.
- Adjust how you use it: query once, early in the task; write specific queries (60+ characters); after an empty result or error, accept "nothing" or fix the error instead of retrying; check whether injected Working Memory is stale.
What the author says
ZinkLu @ZinkLu2 (X article, 12 Nov 2026): of 1,050 recall windows, 101 were confirmed used (lenient 12.4% = 101/814, strict 9.6% = 101/1,050); only 57 both added new information and drove the next step; 159 were empty and 137 errors. Codex recalled most (94 per 100 prompts) but had the lowest adoption, 9.1%; OpenCode searched less and had the highest, 29.1%; Claude Code, Kimi Code and Hermes rarely searched on their own, and although Claude Code's plugin added a reminder to every message (1,619 times), only 11 of 162 sessions searched. Project-management work had 21.7% adoption (issue formats, board conventions) versus 5.1% for research. Memories from the same agent were used 14.7% versus 6.0% from other agents. Injected Working Memory stayed at the 6 August version through September and October. 52 windows showed stale or conflicting information. He stresses that 12.4% does not mean memory helps only 12.4% of the time, that missed recalls weren't counted, and that there's no no-memory control. He is a Nowledge Mem user and ends with a plug for it.
Our test
We don't have his session logs and did not run the same review with Nowledge Mem; all numbers are his. We checked the numerators and denominators in the article (e.g. 101/814, 101/1,050) and the methodology in his appendix.
How to get in
The original is a long Chinese X article with charts: https://x.com/ZinkLu2/status/2120867160428064925 (use browser translate). Nowledge Mem's site and docs are in English: https://mem.nowledge.co/docs .
Who it suits
People who've connected a memory system (Nowledge Mem or another) to several coding agents and want to know if it's worth it; if you don't want to run the analysis, just apply the rules in the last step.
Pitfalls
- "Used" only counts evidence visible inside the window, not better task outcomes, and missed recalls weren't counted.
- All semantic judgments were made by AI with no human line-by-line check; two AI reviewers agreed 91.7% on a sample.
- When re-running old queries, half of today's top-5 results were written after the original query, which can inflate results unless filtered.
- Auto-injected memory can go stale for a long time; his was over two months old.
- Findings come from one person's workload and shouldn't be used to rank agents. We did not reproduce the analysis.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- AI backend code that guards against races, leaks and bad migrationsCodingbackendskillsClaude CodePostgreSQLFree; about 2× tokens per task◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas