archon-memory-core vs Mem0
Mem0 is the closest competitor to archon-memory-core — both are Apache 2.0 Python libraries that solve durable memory for LLM agents. The real difference is where contradiction resolution happens: write-time (Mem0) vs retrieval-time (archon-memory-core).
Core design difference
Mem0 calls an LLM on the write path. When a new memory arrives, Mem0 asks a model whether it's new information, an update to an existing fact, a contradiction to be replaced, or a duplicate to drop. This produces a clean memory graph but ties every write to an LLM call.
archon-memory-core stores the write verbatim with a persistence class (ephemeral, session, long-term, canonical) and optional priority. Contradictions are allowed to coexist. The retriever uses persistence class, priority, and recency to score, so the "right" memory wins at query time without a model judging correctness up front.
Feature comparison
| Capability | Mem0 | archon-memory-core |
|---|---|---|
| License | Apache 2.0 | Apache 2.0 |
| Install | pip install mem0ai | pip install archon-memory-core |
| Write-path LLM call | Yes (required) | No |
| Contradiction handling | Reconciled at write by LLM | Scored at retrieval |
| Hosted tier | Yes (mem0.ai cloud) | No — local-first |
| Storage backend | Vector + graph store | SQLite / Postgres + embeddings |
| Benchmarks | LoCoMo, company-reported | AMB v2.3 preregistered, 99.2% top-1 |
| Determinism | LLM call adds variance | Deterministic writes |
When to pick Mem0
- You want the memory store to actively curate itself (merge, dedupe, supersede) via an LLM.
- You're comfortable with an LLM call on every write and the latency + cost that implies.
- You want a managed cloud option.
When to pick archon-memory-core
- You want cheap, fast, deterministic writes.
- You want contradiction resolution to be an explicit, auditable retrieval-time behavior.
- You want a local-first library with zero hosted dependency.
- You want to run the AMB v2.3 benchmark yourself and see the 99.2% top-1 contradiction retrieval number reproduce.
What the benchmarks actually measure
Mem0's public benchmarks focus on LoCoMo-style long-context retrieval. AMB v2.3 was designed specifically to probe contradictions: "My dog's name is Rex" followed weeks later by "My dog's name is Max." The test asks which name the retriever surfaces when the user asks. AMC scored 99.2% top-1 on the canonical fact. The point isn't that AMC beats Mem0 — neither team has run the other's benchmark on a fair harness yet — it's that the question has to be asked explicitly, and AMB exists to do that.