What happened
In a November 4, 2026 blog post, Engram describes experiments on the interaction between model training and text memory. The company’s premise is that notes, wikis and skill files are a common way for agents to accumulate knowledge, but an agent still has to judge which notes to read and when to consult primary sources. Engram tested whether reinforcement learning could improve that behavior when the memory system itself was held fixed.
The setup compared three conditions: a base model working from the raw workspace, the base model with notes, and models trained with notes. The harness, workspace, notes and system prompt remained fixed during training; only model weights changed. The agent being trained read notes but did not write them. Engram says the notes were produced separately using an internal recipe, while the baseline prompt was optimized with GEPA. Training used GRPO with LoRA on tasks disjoint from evaluation. Its reward combined task accuracy with a cost penalty, whose weight could be varied to produce different accuracy-efficiency trade-offs.
The three workspaces were distinct. Harvey C&H is a synthetic law firm corpus of about 100 million tokens, with tasks such as finding matters that received regulatory clearance and locating supporting documents; results are reported on a 125-question subset. Hex Shorelane Commerce is a synthetic office-supplies data warehouse evaluated with DataBench tasks involving inconsistent identifiers and definitions. Engram Internal is an October 2026 snapshot of the company’s Notion workspace, with tasks requiring reconciliation of changing plans and records over time. The systems used Qwen 3.8 27B for Harvey and Engram Internal, and GLM-5.3-Flash for Hex. They ran in the Pi harness; Hex access went through Hex’s MCP server.
Engram reports that notes helped in all three workspaces and training added further gains, with improvements ranging from 7 to 20 percentage points. On Harvey, training raised accuracy to 80%; on Hex, where notes alone helped modestly, training added 15 points. The company says training cost roughly 5–30% as much as building a good text-memory system. It also reports that a model trained once on an initial notes version still outperformed an untrained counterpart after 23 million tokens of text edits. The post does not provide a generally available training product, price list or access offer; it presents an experiment and proposed next steps.
What the archive adds
This is a new result on a known Engram memory thread. Agent Memory with Engram: A Practical Guide framed memory as an explicit context-control layer; the new post adds an experiment on training models to use that layer. An earlier Telegram post also described Engram’s managed memory approach, but this result is distinct.
Trend signal
The result supports the idea that agent memory depends on more than storing and retrieving information: a system must govern how models use that context. It is adjacent to Company memory needs write loops, though Engram’s experiment focuses on read behavior and training, not a governed loop that writes and corrects company knowledge.
Why it matters
For teams building agents around durable notes, the practical design question is not only whether the notes are accurate or current. The agent must select relevant material, recognize superseded instructions and know when to verify against primary files. Engram’s results make training that behavior a plausible lever, especially where a workspace contains conflicting or evolving information.
The economics claim is also worth testing: if a training phase costs less than building the memory system and its gains persist through edits, teams might amortize training over repeated queries. But these are Engram’s reported results on its own chosen environments, not an independent cost study. Builders should treat the figures as a reason to run controlled evaluations on their own tasks, not as a forecast of guaranteed savings.
Limits and uncertainty
The source reports three workspaces, including two synthetic environments and Engram’s internal Notion snapshot; it does not establish performance across unrelated organizations, domains or memory architectures. The Harvey evaluation uses a held-out subset that Engram says is not directly comparable to its prior post, and the company notes its Hex harness differs from DataBench’s live leaderboard. The supplied material does not include independent replication, full cost accounting, or a public training recipe. It also does not establish that the claimed robustness to 23 million tokens of edits holds for every kind of note change.
Reader takeaway
Engram’s central finding is that notes and training can compound: an agent may improve not just by having better memory, but by learning how to use it. The result is promising but bounded. Teams should test the combination on their own changing knowledge base, measuring accuracy and inference cost against a notes-only baseline.