The Memory Spine: What Running a Company on Gated Agents Taught Us 6 Months Before the Frontier Papers
The Memory Spine: What Running a Company on Gated Agents Taught Us 6 Months Before the Frontier Papers
Subtitle: Yesterday, Anthropic published "The components of agent memory." Here is what happened when we built and operated that same architecture in production, starting April 2026. Author: Abhishek Srivastava Date: November 7, 2026 Updated: November 9, 2026 Canonical: https://eleven11.pro/blog/the-memory-spine
Yesterday, Lance Martin and the engineering team at Anthropic published an insightful research piece: "The components of agent memory".
Their central thesis is sharp and necessary: durable agent memory needs an organized hierarchy, from raw transcripts at the bottom to a small index loaded into every session at the top. The post makes four moves:
Structuring memory as linked markdown wikis rather than flat storage or pure semantic search.
Keeping the always-loaded context small: a lean index that pulls relevant context on demand.
Offline consolidation ("dreaming"): an idle-time process that checks notes against the original transcripts, corrects them, merges duplicates, and builds the wiki.
Mid-task salience: letting strong models judge what is worth remembering, rather than fixing rules in advance for how they record notes.
We read the paper with immense satisfaction and quiet pride.
Because since April 2026 we have been building that same architecture, and today it runs every day. The linked markdown wiki, the token-light grounding snapshot, and the background distillation daemon are the operational heartbeat of Eleven11.
We didn't develop this pattern to publish a benchmark paper on simulated environments. We built it out of sheer necessity in the arena, because running real client workloads and managing complex software fleets across autonomous agents without a persistent memory spine broke down within 48 hours.
Here is what we built, our dated production receipts from April and June 2026, and what running this architecture in the wild teaches you beyond the laboratory.
1. The Immutable Record: April 20 to June 30, 2026
Our memory spine has a birth date. On April 20, 2026, we committed its schema. On June 30, we filed the operating contract every agent session still follows.
Receipt A: The Memory Spine Schema (2026-04-20T13:48)
From commit c56bc70d in our Architect repository, "feat(architect): slice 1 memory spine", migration drizzle/0007_memory_spine.sql:
"CA1-recoded recall index. Every event worth recalling lands here."
The memory_units table it creates gives every memory a saliency score, a replay_count, and a last_replayed_at timestamp. The name comes from the hippocampus, the same neuroscience of memory consolidation (McClelland, McNaughton and O'Reilly, 1995) that Anthropic's post draws on. That commit is six and a half months older than the post.
Receipt B: The Memory Spine Defined (2026-06-30T14:17)
From wiki_claim-a-linked-matter-64f8aa08:
"A matter like jibhi.in / ailyak / clearstate is a CUSTOMER we build tech for... Linking makes that local project folder the synced matter FIRST, before any ask fires; the matter is both working tree and memory spine. Ops knowledge lives in the wiki from now on; Architect surfaces + works + shares it."
Receipt C: The Operating Model (2026-06-30T22:54)
From wiki_claim-architect-operating-model-a836389b:
"1. Scope precedes prompt. Pick a matter via e11-ops BEFORE working. There is no workspace-wide agent. Everything starts with the matter picker; tools, recall, and asks are all matter-scoped.
2. Recall → work → checkpoint. Recall the wiki before re-deriving or contradicting a prior decision; checkpoint what changed when done. The wiki is the shared human+agent memory spine; you deposit into it, you don't own it. Token-light: grep, not transcripts."
These were not theoretical notes in a design scratchpad. They became code, shell scripts, and daemons that have governed every single agent session across our fleet ever since.
2. Comparing the Architectural Pillars
| Anthropic Paper (Nov 6, 2026) | Eleven11 Production Architecture (Built April to August 2026) | The Operational Reality |
|---|---|---|
| Wikis over Flat Storage"Representations like wikis can help Claude traverse its ever-growing knowledge base." | The Typed Markdown Spine (e11-wiki)Every memory unit is a structured Markdown document (checkpoint, itch, wiki_claim, annotation) carrying machine-readable frontmatter (id, supersedes, deprecated, saliency). | Flat logs decay; embeddings lose causal order. A typed wiki with explicit supersedes links allows agents to update beliefs without hallucinating historical revisions. |
| Tiny Always-Loaded Layer"Injected memories use up valuable tokens, and an incorrect note can be carried into every session." | Token-Light Grounding Snapshot (attach.sh)Session initialization injects titles only, zero bodies. Deep memory is loaded on demand via targeted semantic + lexical + graph recall.sh. | Injecting full history poisons the agent's attention. Injecting a 20-line index of open itches and recent decisions grounds the agent in <200 tokens. |
| "Dreaming" / Offline Consolidation"Dreaming gives Claude a way to test its notes against the original transcripts" | The Three-Tier Consolidation Engine1. distill.sh: Sweeps raw agent transcripts into verified claims.2. reconciler: Background daemon managing bidirectional sync and auto-retiring closed itches.3. Gardener: Weekly cron auditing orphan notes and sync conflicts. | Live sessions focus on execution. Consolidation belongs in background daemons running while the team sleeps. |
| Salience & Task Predictions"By writing down a rule, an agent predicts that the pattern it's describing will be relevant again." | The itch ContractAny agent finding friction mid-task deposits an itch-*.md unit with explicit owner tags and target personas. | Don't force an agent to fix everything mid-stride. Let it register friction concisely for the next specialized seat to resolve. |
3. What Laboratory Benchmarks Miss: The Production Frontier
Anthropic evaluated their memory framework on CL-Bench, a corpus of 350,000 internal documents, and the long-horizon game Pokémon Red. Those are excellent testbeds for isolating episodic memory loops.
However, when you apply structured agent memory to real-world engineering, customer data, and commercial systems, three massive architectural challenges emerge that isolated loops never encounter:
A. The Multi-Persona Fleet
Anthropic's experiments focus on a single agent learning from its own past.
In our production fleet, memory is shared across distinct specialized agent personas:
Sthapati (the design layer, architectural rulings, ceremony)
Sutradhaar (the board-drain weaver, build orchestration)
Sakshi (adversarial verification, receipts, Edge-1 testing)
Dhruva (fleet telemetry, background watcher)
An architecture seat cannot have its context cluttered with shell build logs; a verification seat must not be biased by an orchestrator's optimistic assumptions. By sharing one typed markdown memory spine, governed by per-seat identity tokens (E11_OPS_ACTOR) and asynchronous nudge queues (send-nudge), multiple specialized agents collaborate over the same project memory without stepping on each other's toes.
B. The Triangle: Memory Must Be Governed by a Human Gate
If an agent has full autonomy to rewrite its own long-term memory without human oversight, a subtle defect or hallucination will compound across sessions. Anthropic's post shows how this happens: in Pokémon Red, one wrong note about which side of the mountain the PokéCenter stood on sent Claude back down the route it had come from, until dreaming checked the note against the transcript and rewrote it.
In Eleven11, we solved this with The Triangle (Human + Agent + Architect):
Agents propose moves; they never execute irreversible outward actions directly.
All ask-lifecycle transitions, code merges, and major belief graph updates route through a unified human decision surface (
/today).Memory is immutable and auditable. Every decision carries who decided it, when, and against what evidence commit.
C. True Data Sovereignty
Where does the memory live?
If agent memory lives exclusively inside a proprietary vendor's hosted context or closed cloud database, the enterprise customer has effectively surrendered their institutional intelligence to a third party.
Eleven11 was engineered to be sovereign by default:
The memory spine lives on-box in standard Markdown and PostgreSQL/pgvector.
It operates in dual-mode: offline-capable via peer-to-peer file sync (Syncthing) or synchronous HTTPS box-as-truth API.
The frontier LLM (whether Claude, Gemini, or a local open-weights model via Ollama) is treated as a stateless reasoning engine. The memory remains entirely on the owner's sovereign machine.
4. The View from the Arena
There is an old observation in software engineering: if you build at the frontier of real constraints, you will inevitably arrive at the same truths that research labs later formalize.
We welcome Anthropic's paper. It provides rigorous empirical validation for an architectural pattern that many dismissed as "just markdown files" when we started building it in April. It confirms that the future of agentic AI is not bigger prompts or monolithic black-box vectors, but structured, auditable, and human-governed memory spines.
To Lance Martin and the Anthropic research team: thank you for publishing the data. You proved the science.
To the builders in the arena: the architecture works. We have six months of production commits proving it.
The future isn't just agents that remember. It's sovereign fleets that know how to build together.
Written by Abhishek Srivastava, founder of Eleven11.
First drafted and recorded to the Eleven11 memory spine on November 7, 2026. Updated November 9, 2026 to start the record at our April 20 schema commit.