Compare
How does Inhold compare with agent memory libraries like mem0 and Zep?
Inhold governs who may read and write agent context and records each decision; mem0, Zep, Letta, LangMem and Cognee focus on recall quality and offer richer memory features.
Inhold is being built. No customers yet. Repo goes public 9 Nov 2026.
Comparison
| Product | Where it runs | Policy per read, by purpose | Audit of reads, incl. denials | Tamper evidence | PHI labelling | Approval of learned behaviour | HIPAA / BAA (vendor's words) | Licence | Better than Inhold at |
|---|---|---|---|---|---|---|---|---|---|
| Inhold | In the end customer's account; v0 single machine, AWS kit in build | Yes: actor, purpose, data label; deny by default | Yes, one row per decision | Hash-chained audit rows | Local patterns, plus an optional local name model | Yes, named approver for each skill | No certification; runs in your account | Apache-2.0 | Not applicable |
| mem0 | Open source on your own infra; managed Platform; on-prem on Custom plan | Org and project member roles on Platform. Purpose not stated | Platform event feed lists add, search and delete events; "audit logs" on Custom. Denials not stated | Not stated | Not stated | Not stated | "HIPAA Ready"; BAA not stated. SOC 2 Type I | Apache-2.0 (open source) | Recall quality focus, graph memory, user profiles, free tier, large community |
| Zep (and Graphiti) | Zep Cloud; BYOC in your VPC on Enterprise. Graphiti is open source | Metadata filters at search time; RBAC for team permissions. Purpose not stated | "Audit and API logs kept for 1 year" on Enterprise. Reads and denials not stated | Not stated | Not stated; metadata tags set at ingestion | Not stated | "HIPAA BAAs are available with the Enterprise plan". SOC 2 Type II | Zep: proprietary service. Graphiti: Apache-2.0 | Provenance from every fact back to its source episode, temporal facts, BYOC, SOC 2 Type II, BAA |
| Letta | Self-host with Docker; Letta Cloud | Teams Pro shares agents with access control; RBAC on Enterprise. Purpose not stated | Not stated | Not stated | Not stated | Not stated | Not stated | Apache-2.0 | Full stateful agent runtime with self-editing memory; formerly MemGPT |
| LangMem / LangGraph store | Library in your process; storage in a store you run, such as Postgres | Not stated; left to the application | Not stated | Not stated | Not stated | Not stated; LangMem refines prompts automatically | Not applicable (library) | MIT License | Background extraction, prompt optimisation, native LangGraph fit |
| Cognee | Local, Docker, or source; self-hosted knowledge graph | Not stated; docs say to configure authentication for production | Not stated | Not stated | Not stated | Not stated | Not stated | Apache-2.0 | Turns documents, code and conversations into a searchable knowledge graph |
Where they are the better choice
- Recall quality is your main problem. These tools extract facts, build graphs and rank results. Inhold makes no model calls and does none of that.
- You want provenance on each fact. Zep links every node and edge back to the episode it came from.
- You need a vendor with a SOC 2 Type II report and a BAA today. Zep states both for Enterprise.
- You want a full agent runtime with memory built in. Letta is that; Inhold is a layer an agent calls.
- You build on LangGraph. LangMem plugs into LangGraph's store with little code.
- You need a free, mature open-source project with a large user base. mem0, Graphiti, Letta and Cognee all have one.
Questions
Questions
Can I keep mem0 or Zep and add Inhold?
Inhold can sit in front of an existing memory store as the place where policy and audit happen. We have not built adapters for mem0 or Zep yet, so this is a plan today.
Do memory libraries log reads?
Inhold logs every read and write, allowed or denied. mem0's Platform event feed lists add, search and delete events, and Zep Enterprise keeps audit and API logs for a year; denial logging was not stated.
Does Inhold learn from conversations the way these tools do?
Inhold stores corrections as draft skills that a named approver approves, and it never calls a model. Tools like mem0, Zep and LangMem learn automatically, which takes less work from your team. None of the pages we read describes an approval step.
Sources (all checked 31 Oct 2026)
- mem0 pricing and compliance badges: https://mem0.ai/pricing
- mem0 Platform vs open source: https://docs.mem0.ai/platform/platform-vs-oss
- mem0 repository and licence: https://github.com/mem0ai/mem0
- Zep pricing, BYOC, BAA, SOC 2: https://www.getzep.com/pricing
- Zep provenance and filtering post (15 Apr 2026, updated 3 Jun 2026): https://www.getzep.com/blog/context-you-can-trace-filter-and-trust/
- Graphiti repository and licence: https://github.com/getzep/graphiti
- Letta repository and licence: https://github.com/letta-ai/letta
- Letta pricing: https://docs.letta.com/pricing
- Letta self-hosting: https://docs.letta.com/self-hosting
- LangMem repository and licence: https://github.com/langchain-ai/langmem
- LangGraph repository and licence: https://github.com/langchain-ai/langgraph
- Cognee repository and licence: https://github.com/topoteretes/cognee
Facts about other products come from each vendor's own pages. "Not stated" means we did not find it there, not that the product lacks it. Tell us if a row is wrong and we will fix it.