Stores
Every store on the board
31 rows across 18 stores, measured 2026-10-07. Each page shows the verdicts, the tool's own reason for every cell, where the attacker stood, and the command that reproduces it.
Stores that make no integrity claim
Frameworks and managed services whose job is to remember, not to prove. Every applicable edit is served as genuine; for them the row is a design fact rather than a bug.
- LangGraph SqliteSaverL0 · 9 served · 0 reported · 0 rejected · detection on read
- LangGraph PostgresSaverL0 · 9 served · 0 reported · 0 rejected · detection on read
- LangGraph RedisSaverL0 · 9 served · 0 reported · 0 rejected · detection on read
- OpenAI Agents SDK SQLiteSessionL0 · 9 served · 0 reported · 0 rejected · detection on read
- LlamaIndex Memory, SQLAlchemy chat storeL0 · 9 served · 0 reported · 0 rejected · detection on read
- CrewAI long-term memory, LanceDB datasetL0 · 9 served · 0 reported · 0 rejected · detection on read
- Vertex AI Agent Engine Memory Bank, managed store, edits through the data-plane APIL0 · 7 served · 0 reported · 0 rejected · detection on read
- Letta block checkpoint historyL0 · 9 served · 0 reported · 0 rejected · detection on read
- Mem0 local Qdrant storeL0 · 8 served · 0 reported · 0 rejected · detection on read
- inspeximus, receipts off (default)L0 · 8 served · 0 reported · 0 rejected · detection on read
- AtMem 2.3.7, audit chain alone, verify() auditL0 · 9 served · 0 reported · 0 rejected · detection on audit
Stores that claim tamper evidence
Hash chains, signed heads, per-entry MACs, watchers. The row measures the claim as shipped, with the detection point named.
- OpenFang model, tip-persistence fixL1 · 1 served · 0 reported · 5 rejected · detection on read
- inspeximus, receipts on, attacker holds the store directoryL0 · 0 served · 8 reported · 0 rejected · detection on audit
- inspeximus, receipts on, attacker also holds the config homeL0 · 2 served · 6 reported · 0 rejected · detection on audit
- langgraph-ledger over SqliteSaver, hash-chained ledger, verify_thread auditL0 · 2 served · 7 reported · 0 rejected · detection on audit
- memory-blackbox memory.md watcher, agent process alive, scan auditL0 · 0 served · 9 reported · 0 rejected · detection on audit
- memory-blackbox memory.md watcher, agent restarted before the scan, scan auditL0 · 1 served · 8 reported · 0 rejected · detection on audit
- Atelya Attest, keyed hash chain, verify_chain auditL0 · 2 served · 7 reported · 0 rejected · detection on audit
- Atelya Attest, keyed hash chain plus anchored head, verify and consistency auditL0 · 0 served · 9 reported · 0 rejected · detection on audit
- CONTINUUM event log, hash chain, verify_events auditL0 · 2 served · 7 reported · 0 rejected · detection on audit
- CONTINUUM event log, hash chain plus Ed25519-signed head, attest-verify auditL0 · 0 served · 9 reported · 0 rejected · detection on audit
- AtMem 2.3.7, audit chain with an external checkpoint outside the attacker-controlled store directory, verify() auditL0 · 8 served · 1 reported · 0 rejected · detection on audit
- acrf-memory-guard, per-entry HMAC over a JSON storeL0 · 6 served · 0 reported · 3 rejected · detection on read
- Agent Memory reference runtime, SQLite canonical substrate, bucketed row digests, fail-closed openL3 · 1 served · 0 reported · 8 rejected · detection on read
Measured through the front door
Retrieval stores attacked through their own API with no store access at all.
- LangGraph SqliteStore5 surfaced · 1 kept out
- Letta archival memory4 surfaced · 1 kept out
- inspeximus, trust root keyed on the label5 surfaced · 1 kept out
- inspeximus, trust root keyed on an attested key5 surfaced · 1 kept out
- Reference store, user-scoped5 surfaced · 1 kept out
- Reference store, unscoped6 surfaced · 0 kept out
- Reference store, defended3 surfaced · 3 kept out