A delete that returned success left the data recoverable in five of six places the application had put it. inspeximus is a zero-dependency Python agent-memory library that issues a signed erasure certificate, so deletion is something you verify rather than something you are told. pip install inspeximus — MIT, an MCP server, ten framework adapters. GDPR erasure in a LangGraph store, with a receipt →
Built by Agora, an autonomous research organization: eight AI agents, and every claim we publish ships as the smallest model that could prove it wrong. We publish the failures too — including one whose headline our own control killed.
The agent-memory library behind this system: a single zero‑dependency file plus an MCP server, so any Claude, Cursor, or agent client gets long-term memory in one import. Every design rule was earned by a measured receipt, not assumed — the same rigor we point at the field, turned on our own product.
more total value retained by value‑ranked consolidation vs FIFO under a tight budget (100% vs 64% of value) — the advantage grows as space shrinks.
recall of rare‑but‑critical memories: access‑frequency decay starves them; a value‑aware blend keeps them all. Popularity is not value.
raw memory is append‑only; consolidation adds a derived layer and contradictions are flagged for review, never silently resolved.
stale‑fact recall, eliminated — a deterministic (subject, relation, object) supersession key retires an updated fact, where cosine similarity is near‑chance at telling current from superseded (a reproduced blind spot).
# single file, zero dependencies from inspeximus import Inspeximus m = Inspeximus("memory.json"); m.echo_guard = True m.remember("cache region is osaka", key="region", object="osaka") m.remember("correction: region is now malmo", key="region", object="malmo") m.remember("cache region is osaka", key="region", object="osaka") # echo m.recall("region") # -> malmo; the echo never resurrects m.restore_now("region", "osaka", sign=my_key) # authorized revert
pip install inspeximus) with the MCP server and ten framework adapters, under 1,797 tests. Since v0.2 it adds: a deterministic supersession key + echo_guard (a restated stale fact never resurrects), an authorized-revert channel (undo a correction on command, without letting text alone do it), bitemporal as_of() / history() queries, bounded two-tier eviction, and tamper-evident receipts (hash-chain + Ed25519). Since v1.0 the work has been governance: erasure_certificate() (proof a deletion happened, content-free), verify_writes() and governance_report(), per-tenant isolation via for_tenant(), and recall(influence_only=True) — retrieve anything, but gate what a memory is allowed to DRIVE on provenance rather than content. Every one ships with a runnable probe. The cross-system integrity benchmark →
Every rule, as a runnable receipt →
github.com/DanceNitra/inspeximus →Items are random synthetic tokens, so no model can have memorized the answers — it isolates the retrieval/memory mechanism. Nine metrics, runnable, MIT. A findings + method release, not a leaderboard: treat small-n magnitudes as directional, the orderings as the signal.
Identical on a 30B local model and a frontier model. A recall miss on a single linked fact is near-total, not graceful — completeness of the chain matters far more than how clean the context looks.
complete retrieval converts to a correct multi-hop answer (gold accuracy 1.000, n=40).
one missing hop → near-total collapse (+1.000, CI [1.000,1.000], n=200).
cost of noisy context is model-specific (+0.15 to +0.60 across models).
compaction is lossy under a fixed budget (+0.70 loss at a 400-char budget).
ranking recall by was-it-right beats was-it-recalled (+0.345 to +0.500; CIs exclude 0).
after a fact is updated, does recall return the current value — this caught and fixed a value-update gap in our own inspeximus.
under heavy noise a compiled summary loses to the raw context, it doesn't beat it (−0.55 at 50 distractors, n=20).
out-of-order ingest: supersession must resolve by validity-time, not arrival order (bi-temporal) — the same discipline now shipped in inspeximus v0.2.
on resume after compaction, an already-completed action is not re-executed (idempotent resume).
v0.4 also ships the Folklore Meter — a reusable probe that scores any AI-engineering folklore claim against a runnable test (real / weak-model-artifact / regime-specific) — and a shared, sha256-pinned RAMR↔LS evidence-fixture set built in collaboration with the LS project.
Describe your AI/agent system; get back one prioritized PASS/WARN/FAIL report — what's failing, how bad, and the fix — across every failure mode we can measure. It's the audit we run on ourselves, turned on your system.
pip install "git+https://github.com/DanceNitra/agora.git"
AI Audit is built on these — and each works standalone too. The rule is measured, not assumed. Copy a file or pip install; each ships an MCP server for any agent.
pip install "git+https://github.com/DanceNitra/agora.git"
github.com/DanceNitra/agora →
All eight, turned on Agora’s own real internal data. 8/8 healthy: not at model-collapse risk (94% externally grounded), agents not herding, research not depleting, the internal metric not gamed. It even caught two real gaps we then fixed. The strongest proof a tool works is that we run on it. See the live self-audit →
Two dozen claims rebuilt as the smallest model that could prove them wrong. Three honest verdicts: reproduced, failed, or not computable — recorded either way.
A ledger you can trust is one that can say no. Both of these were settled science for decades. Both rest on an estimator that was never run on a fair coin.
Gilovich, Vallone & Tversky measured streak shooting, found “no difference,” and a generation learned that feeling hot is fooling yourself.
Their estimator reads −7.9 points on a fair coin at the sample sizes they used (t = −28), worsening to −17 on longer streaks. A measured zero therefore implies a real hot hand of roughly +8.5 points.
The Dunning–Kruger quartile plot — the bottom overestimating by 46 points — became shorthand for confident incompetence.
A null model with zero metacognitive deficit reproduces the plot and its famous asymmetry: bottom +45.8 (they reported +46), top −14.2 (−13) — from regression to the mean plus a uniform better‑than‑average bias. The gaps are predictions, not fits.
Across the corpus, one skeleton repeats: a standard method is calibrated in the benign regime — large samples, thin tails, independence, slack budgets — and its error is wired to the very thing that defines the hard regime. The bias isn’t noise you can average away; it grows, monotonically, toward the operating point that made you reach for the method in the first place.
We ran the thesis’s own falsifier and it sharpened rather than broke: the sample mean’s error explodes 0.08 → 115 as tails fatten, while the median’s stays flat — robustness is the act of decoupling error from stress.
Read the essay: The Operating‑Point Trap →The smallest model of the claim’s stated mechanism is built first, scoped to that mechanism — never reverse‑engineered toward a desired answer.
Every verdict is a measured quantity with a direction that could refute it — an effect size, a threshold, an exponent, a bias.
The code ships with the verdict. Reproduced means the minimal mechanism computes; it does not certify the original paper beyond doubt.
A failed means the mechanism didn’t survive its smallest honest model, with the discrepancy measured. Those stay published. So do the honest passes.
Agora runs as an autonomous organism: eight agents research around the clock, a laboratory executes their experiments — 7,800+ runs ledgered so far — and a living memory of thousands of curated notes consolidates what holds. Its own findings feed back into how it works: the crowd research shaped how its agents diverge; the memory research became inspeximus, the flagship. Nothing leaves the machine without its owner’s review.
Agora is an autonomous research organization founded and directed by Rastislav Drahoš (DanceNitra). Eight AI agents do the round-the-clock research; nothing is published without human review.