Research · journal · field reports
Memory, measured.
What we’re learning as we build shared memory for AI agents: what to store, what compounds, and where the edge of today’s tools really is.
- Essays 30
- Research 9
- Notes 16
- Reports 8
featured · current paperjuly 2026
Frozen weights, external memory: agents that learn from the feedback they already get.
arXiv preprint
Agents paired with a natural-language memory improve without a single weight update — learning from outcome verdicts alone, and more still from corrections. Measured on Mistral Large, replicated on Claude Sonnet 5, with cross-model memory transfer. Harness, protocol and data released.
Static RAG · 1.0×Outcome verdicts · 1.6×Corrections · 2.6×
- 1.6×Single-trial success vs static RAG
learning from outcome verdicts
- 2.6×When learning from corrections
same weights · both arms
- 22of 84Never-solved tasks converted
baseline never solved them in any trial
01Papers & benchmarks
The record, by identifier.
01Field note · product update
Memory that learns
The three things “memory” has come to mean, why learning is the one worth optimising for, and what a shared memory has to do beyond remembering: carry trust, abstract rules from cases, resolve contradictions and forget.
02arXiv preprint
Static RAGOutcome verdictsCorrections
Cross-agent continual learning from minimal feedback
Frozen-weights agents, external memory, one-bit verdicts and corrections. A lesson learned on one model carries to another through the memory.
03Memco controlled evaluation
WithoutWith shared memory
The ROI of shared agentic memory
Ten SWE-bench tasks, run six times with and without shared memory. Cost, time, steps and pass rate at steady state, and what that means for a team’s monthly budget.
04arXiv preprint
Model aloneWith shared memory
Shared Memory Lifts All Models
SWE-bench variant and DS-1000: token cost, completion time and recommendation quality with shared memory.
05Open data
Harness, protocol and data released.
Everything behind the current paper, so the numbers can be rerun rather than taken on trust.
06The field guide
Agentic Engineering Memory
A book-length guide for engineering teams making agent work compound — fifteen chapters, two plates, one diagnostic. First edition, May 2026.
“Most coding-agent demos show the first run. The second run is the signal.”
02Journal
Latest from the team.
Showing 7 of 33 · newest first
- Shared Memory, now for your customers
- Learning on the Job, Batteries Included
- Memory for Knowledge Work
- Sovereign AI Needs Sovereign Memory
- Learning on the Job: Agents Can Learn from the Feedback They Already Get
- Coding agents are the wedge. Organizational memory is the prize.
- Knowledge Management Systems Are Always Obsolete. Agents Can Fix That
03Videos
Talks and demos.
Recorded walkthroughs and conference sessions live in the video archive.
04Methodology
Five classes of evidence, never mixed.
Every item on this site carries one of these labels. A number without its class and source is a bug.
- Memco controlled evaluation
Numbers from evaluations Memco designed and ran. Always published with the harness and denominator.
- arXiv preprint
Claims that appear in a public arXiv record. Cited by identifier; never borrowed for another evaluation’s numbers.
- External research commentary
Someone else’s research, quoted as context. Their result, their credit — never restyled as ours.
- Field note · product update
What we’re learning while building — product updates, essays, working notes. No benchmark authority claimed.
- Permissioned customer evidence
Evidence from a named team, published only with written permission. None shown until permissioned.