How is flmnt different from vector-recall memory tools?

Recall tools remember what was said; flmnt knows what is currently true. When a decision is superseded, recall still serves the replaced version because the words look current. flmnt follows the recorded causal relationship to the current decision.

Does flmnt match recall tools on retrieval?

flmnt is equal to recall tools on retrieval and superior on currency: it never serves a stale decision as current, because a superseding edge is surfaced with the decision it replaces.

Why flmnt Proof How it works Economics Compare Your agents

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