01 / 30
Memory Layer · A Practice Area within Cone Red's AI-First Transformation Practice · 2026

Memory,
engineered.

The cognitive substrate for AI agents that remember — across sessions, across languages, across years.

Cone Red AI
The AI-first architecture firm
31 slides · ≈ 18 min · skim in 5
For CTOs · 2026-04-27

Not chat history with retrieval.

Architecture.

An 18-minute deck. Five slides if you're benchmarking.

SLIDE 05
Logging vs architecture
The category-error frame, in two sentences.
SLIDE 13
What we rejected
The Mem0 / Letta / Zep comparison table.
SLIDE 25
Phase status (locked eval)
18→20/20, shadow-run live.
SLIDE 29
"Memory is the moat"
The thesis after frontier-model parity.
SLIDE 30
Two paths to a fit call
High-intent + diagnostic. First slot usually within 48 hrs.

Send this deck to a CTO with 5 minutes. They'll know whether to take the call.

The cost of getting
memory wrong.

A finance director asks an AI assistant about Q2 covenants. The assistant invents a number. The director acts on it. Three weeks later, the audit finds the gap. Two analyst-weeks to reconstruct.

A CTO asks the same question in Russian, gets a degraded answer, switches back to English, and the team quietly stops trusting the tool. The cost isn't the wrong answer — it's the lost leverage.

A regulator asks for the trail of how a decision was reached. The chat log exists. The reasoning doesn't.

Every "AI memory" feature on the market treats memory as a logging concern.

We treat it as architecture.

Tom Shea, CEO OneStream, Q3 FY25 SEC 8-K filing:
"80% accurate is 0% useful for finance."
We treat memory the same way. Locked eval, regression-gated, audit trail.

A Tuesday morning,
in the agent's voice.

Dima opens Telegram and types "where did we land on the targetheart audit."

I don't search. The wake-up loaded the right wing the moment the session started — ~170 tokens: identity, this week's engagements, three open approvals.

From his ping to my first useful sentence:
under three seconds.

A message in Russian — "помнишь, Миша говорил про C12?" I pulled the right Mike. Captured 23 days earlier in a voice transcript. An English-only memory would have lost the thread.

Yesterday Dima contradicted a position from March. I didn't correct him in chat. I flagged it once: "you said the opposite on March 14, want to reconcile or supersede?"

No nagging. No second flag.

Thesis.

"AI memory" is a category error.

What every consumer product ships is chat history with retrieval. The model logs your conversations, embeds them, pulls back the closest match.

That is not memory.
That is grep with a thesaurus.

"Memory is an
architectural decision —
not a feature toggle."

What we evaluated,
and what we rejected.

Eight alternatives you've already been pitched.

The case for not buying any of them.

Option Why we passed
ChatGPT / Claude built-in Locked to vendor. No graph layer. No audit trail. Cannot run on premise.
Generic RAG over docs Static. Doesn't form new memories from live conversation. No decay. No contradiction handling.
Frontier-model long context Per-token cost balloons. Latency degrades past 100K. No persistence between sessions.
MemPalace OSS library Strong metaphor. English-leaning embeddings, single-store, no live formation. We adopted the metaphor and replaced the substrate.
Mem0 / agent-memory frameworks Per-message LLM extraction by default — economics break above moderate volume. Single retrieval branch. No locked eval methodology. Production evidence: 97.8% of 10,134 stored memories judged junk over 32 days (mem0 #4573, public). Without a write-side filter, signal-to-noise collapses within a month.
Letta (ex-MemGPT) Memory-as-feature inside an agent runtime, not a substrate underneath one. Couples you to Letta's runtime.
Zep / managed memory svc Strongest commercial alternative on graph + temporal. Vendor owns the eval methodology — buyer cannot lock their own. Sovereignty gap is structural.
"Build it from scratch" The "I don't know what to evaluate" trap. ~33-hour estimate. Never shipped. We treated it as input, not output.

We didn't pick one. We synthesized. Took the strongest mental model in the space — rebuilt the substrate underneath.

Architecture.

Five layers, because one isn't enough.

Five layers — recency × reinforcement.

L0
Identity
Always loaded · ~50 tokens
L1
Critical facts
Always loaded · ~120 tokens · this week's priorities
L1.5
Structure
Auto-discovered graph · communities found, not hand-coded
L2
Wings & rooms
On demand · seeded by L1.5
L3
Hybrid recall
Three branches in parallel — semantic + lexical + graph — fused
L4
Archive
Decayed memories · queryable on explicit request

Session-start budget: under 200 tokens. RAG eats 4–10K.

Three tiers. Most products run only one.

Tier 0 · free
Every message

Open-source NER + heuristics. Catches ⅔ of memorable content at zero LLM cost.

Tier 1 · cheap
Event-driven

Fires when importance crosses a threshold. Fast model extracts structured facts with provenance.

Tier 2 · deep
Nightly

Stronger model dedupes, supersedes, detects contradictions, refreshes the graph.

Per-seat economics scale with conversation signal, not volume.

Five scars.

The hard problems we already paid for.

  1. Semantic similarity collapses on rare terms. Phase 0: 18/20 → Phase 1a multilingual swap: 20/20.
  2. Single-store retrieval is a trap. Vector-only misses exact phrases. Add lexical + graph + RRF fusion.
  3. Per-message LLM extraction is unaffordable. Free heuristic gate first. LLM only on signal-crossings.
  4. Hand-crafted folder trees don't scale. Auto-discover structure from the corpus.
  5. "Forever" memory becomes a junk drawer. Importance × access × recency decay.

Each one is roughly two weeks of work to discover, scope, and fix when you encounter it cold. We already paid that tab.

Three bets won.

Reproducible from a locked eval set.

Bet 1
Multilingual works in production

Mixed traces — EN question, RU follow-up, UK voice — through the full pipeline.

18/20 → 20/20
same questions, same week
Bet 2
Hybrid beats single-source

Storage substrate joins semantic + lexical + graph branches.

20/20
higher confidence on hits
Bet 3
Live formation works

Tier 0 + Tier 1 wired end-to-end. Real operator traffic, not benchmark.

5/8 facts pulled correctly

The fact of a locked eval set is the credibility signal. Most products in this space cannot show one.

What changes daily.
What you own.

  • Cold pickup <200 tokens. No re-priming. The 4 AM ping feels like resuming.
  • Cross-language continuity. RU fact on Monday surfaces in EN query on Thursday.
  • Quiet contradiction detection. Flag once. Newer wins by default. No nagging.
  • Decay-aware recall. Stale facts stop competing with this-week priorities.
  • Graph-aware entity resolution. "Mike said about C12" → the right Mike.
  • Audit trail. Every fact knows when, where, why, and at what confidence.

Sovereignty is the close.

Your data

All of it. Self-hosted. Your infrastructure, your database, your backups.

Your audit trail

Every fact, every timestamp, every confidence score, queryable on your side.

The kill switch

A single env var flips the whole system off. Zero downtime. Agent reverts to vendor-default memory.

The exit

DB dump + documented schema. We don't hold your memory hostage.

Proof.

The eval is locked. Score regressions block phases.

Phase status.

Phase
Status
Result
0 — Sandbox baseline
Complete
18/20 on locked set
0.5 — Structure layer
Complete
9 communities auto-discovered
1a — Multilingual swap
Complete
20/20; RU failures fixed
1b — Storage substrate
Complete
20/20; higher confidence
2 — Formation pipeline
Complete
Live extraction working
2-shadow — Live integration
Active
Observation phase
3 — Hybrid retrieval (RRF)
Scheduled
Lexical + graph alongside semantic
4 — Decay + contradictions
Scheduled
Tier 2 + L4 archive sweep
5 — Operator cutover
Scheduled
Feature-flagged rollout

What this is NOT.

Anti-overclaim. The credibility bank.

  • Not a replacement for RAG over a doc corpus. If you need to query a 50,000-page contract library, buy a doc-retrieval product.
  • Not a system of record. Doesn't replace the CRM, warehouse, or ledger.
  • Not GDPR-certified out of the box. Compliance is deployment-specific work. Anyone claiming "compliant by default" is selling vapor.
  • Not a drop-in for non-conversational agents. Plug it into a batch ETL pipeline, it works badly.
  • Not a hosted SaaS. This is integration measured in weeks, not minutes.
  • Not "your agent is now sentient." It remembers. It does not reason about its own remembering.

If any of those is what you need, we'll tell you on the first call and send you somewhere honest.

Closing.

Frontier models are at parity by 2027.

Memory
is the moat.

The differentiator is whose agent has a relationship with the operator that compounds across every conversation, in every language, over years.

Bring one example of a memory failure that has cost your team time.

We'll work outward from there.

Path A · high intent
45-minute fit call · this week

Your corpus, language footprint, agent runtime. We tell you honestly if it's a fit. First slots usually within 48 hrs.

Book direct → · dima@cone.red

Path B · category-shopping
Memory Readiness Diagnostic

Six questions, five minutes. Tells you which of the five scars your stack is paying for. Output is yours; nothing leaves your browser.

Take the diagnostic →

A Practice Area within the Cone Red AI-First Transformation Practice · 2026-04-27