Executive brief · grounded in the "Bento Under Test" evaluation

Bento — the memory layer that makes AI agents dependable

An open-source, self-hostable memory & retrieval layer for AI agents. It makes them stay current when data changes, cost ~40% less on repeat work, and stop repeating mistakes — proven with executable graders, losses published, not hidden.

Bento open source · self-hostable Runs on plain PostgreSQL Evidence executable graders · reproducible · 95% CIs

Proof at a glance — all measured, all reproducible

+0.31
Stays current vs a stale cache
accuracy after updates
−42%
Tokens at equal accuracy
−34% latency too
0.46→0.78
Correctness on conflicting docs
+0.32, CIs separate
2.4×
Whole-function code hits
vs fixed-size chunking
75% less
Repeated failed actions
procedural memory
tie/win
Backend vs mem0, every test
0.893 ≈ 0.890

What it saves — measured, then modeled

Cost per recurring multi-step task · at equal accuracy (1.00)

Tokens · baseline
1,767
Tokens · bento
1,032 · −42%
Latency · baseline
49.3s
Latency · bento
32.3s · −34%
Model calls · base
177
Model calls · bento
108 · −39%

From wrong / stale → right / current

Correct? · no memory
0.46
Correct? · bento
0.78
Current? · stale cache
0.61
Current? · bento
1.00

On conflicting docs, memory takes accuracy 0.46→0.78 (+0.32). After an update, bento serves the current answer (1.00) where a non-invalidating cache goes stale (0.61).

Estimated cost savings — apply the measured −42% to your spend

Monthly inference spend on applicable workload*Token/compute savedEst. saving / monthEst. saving / year
$10,000−42%$4,200$50,400
$50,000−42%$21,000$252,000
$200,000−42%$84,000$1,008,000

*Illustrative, transparent math: bento cut tokens 42% (and latency 34%) at equal accuracy on a recurring multi-step workload (measured). The table applies that 42% to the portion of your monthly agent-inference spend that is recurring/multi-step — where bento's recall applies. Single-shot or novel queries see less; your savings scale with how repetitive your traffic is. Not counted (additional upside): −34% latency → fewer SLA breaches, better UX, more throughput per GPU; and avoided wrong-answer / stale-answer incidents.

Tailored to the seat

CEO"Why does this win a market?"

Memory is the layer that turns AI demos into products

Every company is shipping agents; the demo always works, production goes stale, expensive, and wrong. Bento fixes those three — open source (adoption flywheel, no lock-in objection) with a hosted option for revenue.

  • Wins or ties the backend in every mem0 head-to-head.
  • The one cell a competitor "led" was their agent loop, not their memory — bento ties it in the identical loop.
  • We publish where bento loses. In an "AI slop" market, that honesty is the brand.
Moat: reproducible, open-source, honest — what skeptical enterprise buyers trust.
CTO"Is it sound, and do I trust the numbers?"

One API, four ingredients, plain Postgres — auditable end to end

A FastAPI layer composing graph retrieval, supersession-backed currency, code-aware chunking, and extraction — on PostgreSQL + Valkey. No new vector DB to operate. Open source, so your team can audit the exact code behind every claim; self-hostable, so data stays in your boundary.

  • Executable graders (not an LLM judge), 95% CIs, a Threats-to-Validity section.
  • Graph is an enhancer on top of vector+BM25 — degrades gracefully.
  • Stated limits, tested at scale and reported: version replay now delivered (a18); graph ties keyword search on simple lookups; cross-session memory retirement depends on the distill model — a temp-0 instruct model takes stale retention 1.0→0.1, a coder model leaves it at 1.0.
Proof: 0.46→0.78 correctness (CIs separate) · −42% tokens at equal accuracy · self-recovery loop saves ~27 ms / prevented call.
CPO"What customer pain does it remove?"

Memory is the feature that makes agents feel reliable

Each common production complaint maps to a measured fix (table below). Honest maturity line so you can plan against reality:

  • Production-ready: currency, cost-reuse, correctness, code chunking, graph retrieval, point-in-time version history ("what did the KB say in March?", a18 sync), and the autonomous failure-recovery loop (Tier-1, measured).
  • In progress: the loop's optional LLM-analyzer tier (Tier-2).
Drop-in: one API behind your existing agent — no re-platforming.
CRO"Who buys, and how do I handle objections?"

Sell to teams stuck at "the demo worked, production didn't"

Support, ops/SRE, research, and dev-tools teams paying for token bills, latency SLAs, and wrong-answer incidents. The lines that close:

  • Bill: "~40% fewer tokens, ~34% faster, same accuracy."
  • Trust: "won't serve last month's answer."
  • Competitive: "ties/beats mem0; their recall costs ~2× the calls and false-recalls at the frontier (0.42)."
Objections: Hype? → run it yourself. Lock-in? → open source, self-host. Beats mem0? → backend yes.

The complaint → the fix (CPO / CRO detail)

The complaint you hearWhat bento doesMeasured
"Our assistant quoted last quarter's price."Update the source; bento serves the new answer. A naive cache keeps answering the old one.+0.31 vs stale cache
"Our research agent is slow and burns tokens on repeat questions."Recalls the resolved answer instead of redoing the multi-step lookup.−34% latency · −42% tokens
"Support gives wrong answers when our docs contradict."Accumulates verified facts and trusts them over noisy docs.0.46 → 0.78 (+0.32)
"Our coding copilot returns half a function."Code-aware chunking keeps whole functions intact.2.4× complete top-hits
"Our ops agent keeps making the same blocked call."Remembers the recovery procedure and applies it.−75% repeated failures
"It can't answer 'who owns the service X depends on'."Graph traversal connects facts across documents.wins multi-hop where vector/keyword can't

Use cases — where bento fits

Match the agent to the capability it needs. Every row maps to a measured result above.

Agent / workloadPrimary capabilityMeasured
Support & RAG assistantsCurrency + correctness on messy docs+0.31 vs stale cache · 0.46→0.78
Research / analyst agentsCost-reuse on recurring multi-hop questions−42% tokens / −34% latency
Ops / SRE / incident agentsProcedural memory + the self-recovery loop−75% repeated failed ops
Developer-tool / coding copilotsCode-aware chunking + self-recovery loop2.4× whole-function hits
Knowledge-base / compliancePoint-in-time version replayas_of returns the historical version
Multi-document reasoningGraph retrievalwins multi-hop with no lexical bridge

Common deployment patterns

Self-hosted, single-tenant (Dedicated). The whole stack runs on your own PostgreSQL + Valkey, inside your boundary — data never leaves. The default for regulated buyers.

Managed / hosted (Shared). We run it; you call one API. Fastest path to a pilot, no ops burden.

Drop-in behind one API. No new vector database to stand up — bento composes its four ingredients behind a single backend-for-frontend, so you add memory to an existing agent without re-architecting.

Distill model is a deployment decision. For cross-session currency, point the distill step at a capable temperature-0 instruct model, not a coder model — that single choice moved stale-sibling retention from 1.0 to 0.1 in our test.

Real-world: what production looks like

Illustrative end-to-end scenarios (composite, not named customers) — each step is a capability measured above.

Support agent on a live pricing/policy KB. Prices update mid-day → the agent serves the new number, not last quarter's (a non-invalidating cache is wrong about a third of the time). Recurring tier-1 questions resolve from one recalled answer (−42% tokens). When two source docs disagree, accumulated verified facts win (0.46→0.78).

Ops / coding agent in CI. A session hits python3 script.pyModuleNotFoundError, recovers with uv run python script.py. The loop records it once; every later session reads the correction and skips the doomed call — ~27 ms plus a wasted model turn saved each time, 75% fewer repeated blocked operations across the fleet. Corrections that would cost more context than they save are dropped, so the always-loaded block stays small.

Compliance review. "Show me what our retention policy said on March 1" → an as_of query returns the version current on that date, not today's.

The cross-cutting differentiator: the honesty is the moat. This is a market drowning in unverifiable "AI" claims. Bento ships measured numbers, open code, reproducible runs, and a published list of its own weaknesses. That is exactly what a serious buyer looks for and rarely finds.