What changed
Persistent, default-on agent memory (CLAUDE.md files, memory directories, decision logs) became mainstream in coding harnesses within the last ~year, so the first cohort of 6-12-month-old, never-curated memory corpora is now old enough to have accumulated rot (FACT from convergence description; the underlying signals 1424/1419 were referenced but not included in this input).
Why now
Creation of memory was democratized but curation was not; the first generation of aged corpora is hitting quality problems now, and harness vendors have not yet shipped built-in hygiene — a window that will not stay open long (HYPOTHESIS per the convergence's own falsification condition).
Converging signals
The convergence cites signals 1424 (teams experiencing memory drift/sync conflict) and 1419 (loss of oversight), but neither signal's text or URL was provided in this input, so their strength cannot be verified here. No demand_evidence items were supplied.
Customer pain
Agent output quality silently degrades because persisted memory contains wrong facts, obsolete constraints, contradictions, and references to files/flags that no longer exist; owners misattribute the degradation to the model itself and cannot diagnose it (INFERENCE, explicitly labeled as such in the source).
Who pays
HYPOTHESIS: engineering leads at small dev shops, AI-forward agencies, and startups whose coding agents produce client deliverables or production code — the 'money at stake' filter. Reachable via HN/X/dev communities. Willingness to pay is unproven: zero demand_evidence was provided.
Solved today
Manually, rarely: someone occasionally reads and prunes CLAUDE.md by hand, or the team deletes memory wholesale and starts over. Mostly it is not solved at all because owners don't realize memory is the cause (INFERENCE).
Why current solutions are bad
Manual review doesn't scale past a few files, can't cross-check claims against the current repo state (does that flag still exist?), and wholesale deletion destroys the valuable memory along with the rot. Nothing today produces a diagnosis linking a specific stale fact to observed agent misbehavior.
Proposed product
A memory-audit tool: point it at a repo/org, it ingests all agent-memory artifacts, cross-references them against the live codebase (stale file/flag references), detects internal contradictions and duplicates, scans for leaked secrets, and outputs (a) a graded rot report, (b) a proposed cleaned corpus as a reviewable diff. Sell the audit per incident; upsell a scheduled re-scan (CI job) as recurring revenue.
MVP version
A CLI/script that takes a directory of CLAUDE.md/memory files plus the repo, runs staleness checks (referenced paths/flags/deps that no longer exist), secret scanning, duplicate detection, and LLM-driven contradiction detection, and emits a markdown report + cleanup diff. Buildable solo in 2-4 weeks with the founder's existing AI-workflow skills.
30-day build
Run the convergence's own cheap test before building for sale: recruit ~20 volunteered real corpora via HN/X, run the analyzer, publish anonymized findings ('we audited 20 teams' agent memories; here's the rot rate'). Success gate: ≥12 corpora with ≥3 material defects AND ≥2 unsolicited offers to pay.
60-day build
If gated through: productize as a paid audit ($199-$499 per incident) with a free teaser scan (rot score only, findings paywalled). Land the first 5 paying audits from the volunteer/HN audience. Add the CI re-scan subscription ($29-$99/mo) for converts.
90-day revenue plan
Target 10-20 paid audits plus 5-10 monitoring subscriptions — roughly $3k-$8k cumulative. Modest, but the real prize is owning the 'agent memory hygiene' category label early. If the 30-day gate fails, kill at near-zero sunk cost.
Distribution path
Demonstrated value, not relationship sales — exactly the founder's style: the free teaser scan IS the marketing (a shocking rot score on your own corpus), amplified by HN/X show-posts and the published audit study. No ad spend required.
Pricing hypothesis
Per-incident audit $199-$499 depending on corpus size; recurring CI hygiene monitor $29-$99/mo. Per-incident matches how the pain presents (episodic: 'the agent got dumber this month').
Technical difficulty
Low-moderate. Staleness and secret checks are deterministic; contradiction detection is an LLM pipeline the founder already knows how to build. Hardest part is precision — false-positive 'contradictions' would destroy credibility.
Legal / regulatory risk
Low, with one real handling burden: customers' memory corpora contain proprietary code context and possibly secrets, so the tool should run locally/on-prem by default. No regulatory exposure.
Platform dependency
HIGH and this is the core risk: Anthropic (Claude Code), Cursor, and other harness vendors are the natural owners of memory hygiene and could ship it as a built-in feature in one release, per the convergence's own falsification condition. This is a window product unless it expands to cross-harness, org-wide memory governance.
Founder fit
Moderate. Fits his AI-workflow, complaint-mining, and fast-prototyping strengths and his micro-SaaS/report-product preference, and he uses these exact tools daily (he can dogfood on his own server's CLAUDE.md files). But it is NOT his proven government-portal/forced-buyer shape — no mandate, no deadline, demand must be evangelized. Applying the 0.8 gov-portal lesson: this scores structurally lower than mandate plays.
Breakout potential
Moderate: if agent memory becomes standard org infrastructure, hygiene/governance could grow into a real category ('linters for agent context'), with the audit as the wedge. But an incumbent-shipped feature caps the ceiling.
Final recommendation
CONDITIONAL GO — but only on the cheap validation path, not a product build. The 30-day volunteer-audit test costs almost nothing, doubles as the marketing asset, and has explicit pass/fail gates (≥12/20 rotten corpora, ≥2 payment offers). Do not build the productized version until that gate passes, and watch Claude Code/Cursor release notes weekly for vendor absorption. This is a decent window play, not a durable business, and it ranks below any available forced-buyer/government-portal opportunity in this founder's queue.
Next action
Post the recruiting call on HN/X this week offering a free memory audit to 20 teams; simultaneously build the analyzer CLI (staleness + secrets + contradiction passes) against his own server's CLAUDE.md corpus as test data.