The AI memory tool that didn't exist.
Born to be audited.
Just ask.
If there's no memory, luplo says it doesn't know.
Every access, every block, every reason. Forever. Hash-linked and tamper-evident.
The problem nobody calls a problem
Your AI memory is leaving the building.
Nobody calls it a leak.
Who accessed what knowledge and what decisions, and why — nobody knows. luplo starts here.
A departing engineer zips the md folder. And walks out.
Inside: three years of design decisions, meeting notes, incident post-mortems, the "why" behind your architecture.
mem0 / Cognee DB exports into another agent
Memory is structured org knowledge. One command exports it, one command imports it elsewhere — same productivity, no friction. Which decision it backed, what answer it produced — nothing is logged, anywhere.
"Why did we give that answer on October 14?"
A PM asks. Embedding models shift over time. Vector indexes rebuild with subtle differences every time.
By design
Three choices
that close the replay gap.
On the same knowledge, the same query always returns the same result.
Not "looks similar" from a cosine. Because tsquery returns the same output for the same input, and memories are append-only.
Every query carries a tag.
actor_id, timestamp, query, result_ids, prev_hash. Six months later, re-run the audit with the same migration_id — identical output.
Synonyms are resolved by a glossary, not an embedding.
Abbreviations and aliases are expanded via a glossary extracted from the decision corpus itself — not an opaque vector space. Every expansion is auditable: which term, which alias, which decision it came from, and who registered it.
Every access, every block, every reason — forever. Hash-linked, tamper-evident, exportable as SOC 2 evidence.
Why "similar" isn't good enough
"auth" means three things.
Vector-based search treats them as one.
In your codebase: authn (who you are), authz (what you can do), and "auth" as shorthand for either. In a SOC 2 audit they're separate controls (CC6.1 vs CC6.3). In a vector store, they converge into the same cluster.
query: "authz policy"
returns: 12 rows — login flows, password resets, OAuth, 2FA, …
reason: "auth" dominates the embedding
authn / authz collapse into the same cluster
audit: opaque — which rows are authz? which aren't?query: "authz policy"
glossary: authz ≠ authn (conf=0.97)
expanded: (authz | "authorization") & policy & !authn
returns: 3 rows — authorization policy decisions
audit: term #authz sourced from
D-2025-03 "split authn/authz for SOC 2 CC6.1"Grounded in requirements engineering research
luplo's glossary builds on the REGICE / FastText-clustering / ILLOD+ line of work on automated term consolidation (Arora 2017, Bhatia 2020, Hasso 2022) — which establishes that controlled vocabulary with feature-based matching outperforms vector embeddings for abbreviations and spelling variants. luplo extends this to append-only decision corpora. Each time a decision supersedes another, the glossary itself changes — and the glossary becomes both the audit subject and its evidence.
Preemptive control
What you can't access, you can't query.
Draw boundaries with permissions. This file is the policy. luplo filters requests before the agent sees any content. luplo cloud also exposes per-user policy in the dashboard.
# deny-by-default for secrets; log every denial
rules:
- match:
tags: [prod, secrets]
readable_by: [group:sre]
audit_denials: true
- match:
item_type: decision
visibility: internal
readable_by: [role:employee]
deny_export: trueRetrieval in vector stores is vector math. Retrieval in luplo is a permissioned read. Denials leave audit evidence (countable, exportable, defensible).
Where we draw the line
Not what luplo is for.
luplo is not a replacement for vector memory. If your agent needs semantic recall across unstructured chat history, tools like mem0, Cognee, LangMem do that well. Use them.
luplo's lane is different — for the organizational decisions, policies, and facts that must be retrieved the same way every time, by the same query, with a tag.
Honest comparison
Features our peers don't call features.
Every memory product offers recall. The difference is how they recall. They compete on recall quality, embedding cost, integration count. Provability isn't even on their roadmap — that's where the gap opens.
| CAPABILITY |  | vector-based (mem0, LangMem) | graph-based (Cognee, Zep) |
|---|---|---|---|
| Recall mode how results are produced | Deterministic | Probabilistic | Hybrid |
| Same query, 2 mo. later = same result? byte-identical replay | Yes | No | No |
| Synonym / alias handling auditable expansion | Glossary layer | Vector space (opaque) | Entity linking (opaque) |
| Per-query retrieval log actor · timestamp · query · result_ids | Full | Optional | Per product |
| Hash-chained audit tamper-evident line linking | Yes | — | — |
| Policy gate where access control happens | Before read (YAML) | App layer (build it yourself) | App layer (build it yourself) |
| OSS license copyleft vs permissive | AGPL | Apache / MIT | Apache |
Comparison reflects public docs as of 2026-Q2. We'll open a PR if any row is wrong.
One product, different uses
Feel the calm
of deterministic results.
"Every action has a reason, and every reason can be replayed. Born to be audited!"
Hash-chained audit*. Per-actor retention. Tamper-evident exports. Formats auditors already know — a product born to be audited.
"It sits on top of the stack you already have."
No new vendor. No new DB. No new hire. If you're already running Postgres, luplo's operating cost trends to zero.
"Building your own knowledge base? One Postgres will do."
No external embedding API. No vector DB to run. One docker compose up — MCP server included.
"A spec drafted with Claude Code, handed to a teammate using Codex."
Why you chose this approach, why you didn't choose the alternatives, which external sources you consulted — all of it carries through.
"Who read what,
and under whose policy."
Every read carries actor, tag, policy match. DSR, retention. Append-only meets right-to-be-forgotten through tombstones and access blocks.
"Deterministic queries — the trust of an agent that can say 'I don't know'."
For chatbots that must follow a script, vector-based knowledge only piles on uncertainty. luplo brings calm.
* planned for v1.0
Born to be audited.
Ready to be installed.
AGPL core. Self-host with one docker compose up, or try luplo cloud for free.