How flmnt compares · line by line

Equal on recall. Superior on what's true.

Every tool here does recall well — remembering what was said was solved. The lines below are about currency: knowing what's true now, and why. Where a claim has a number, it traces to our public benchmark; run it yourself and check.

capability
flmnt
reasoning layer
Mem0
extraction memory
Zep
temporal graph
Letta
agent framework
DIY RAG
vector store + embed
Where memory comes from
How memories are created
Authored by your agents as they work
Extracted by a second model
Extracted into a graph
Agent-managed notes
Chunked & embedded
Decisions, causes & mistakes as first-class types
partial
entities & relations
Nothing guessed after the fact
partial
agent-written
Currency — what's true now
Serves the current decision after a reversal
partial
time-stamped facts
Supersession test (no textual tell)
1.00
0.00
Reversed decisions served as history, never truth
Causal lineage — the why behind each decision
partial
graph edges
Tracks whether mistakes stop repeating
Recall — what was said
Semantic recall over history
Session-to-session persistence
Economics
Cost per query as the corpus grows
Flat — $0.0077 measured
Grows with corpus
Grows with graph
Grows with context
Grows with corpus
Live ROI ledger, per deployment
Trust & fit
Public benchmark, run every release
partial
published evals
MCP-native — no SDK, no rewrite
SDK
SDK / API
Framework adoption
You build it
Immutable, replayable record
partial
Competitor entries reflect each tool's published approach as of July 2026 — corrections welcome. "Supersession test" is our public benchmark's hardest pack: a decision replaced with no textual tell. flmnt 1.00 · vector recall 0.00, on a frontier model and a small one.
Create your account →See the proof