The model is the brain, REM is the notebook, and the environment is the world the agent acts on. The brain chooses the next move; the world returns files, tool results, and scores; REM decides which notes stay on the brain's small desk.
A naive agent loop puts every observation and answer back into the next prompt, so injected context compounds until the window blows. REM files bulky history outside the model and injects a bounded, useful slice each step: the task can keep growing while the working context stays flat.
|
Illustrative schematic, not a benchmark run. The labels show the shape of compounding versus bounded context.
Status: The bounded read-side win is shipped. Project direction is now context budget over agent horizon; the agent-horizon harness is the next milestone. No horizon multiplier has been measured yet.
REM = Resident Externalized Memory. It is a local-agent memory sidecar that keeps working context bounded and moves background compaction to the AMD XDNA2 NPU, leaving the iGPU available for the foreground model. REM is both a practical tool and a learning repo; start with TEACHING.md for the concepts or the water-audit worked example for one task from start to graceful handoff.
- The compactor keeps up in the committed probe. It drains at ~73 tok/s while the example transcript grows at ~30 tok/s, a ~2.5× keep-up margin.
- Bounded assembly reduces prompt work. RSB-3 fell 47.6%, from 27,386 to 14,341 tokens. Prefill changed from 34.16 s to 0.12 s; after the 30.27 s re-prefill tax, the net saving was +3.77 s.
- NPU placement has a measured contention budget. Concurrent compaction cost ~3.8% iGPU decode versus ~4.8% for the CPU control (3 runs × N=20).
- The context wall is real. Naive assembly overflowed at 37k–58k tokens
against a ~32–40k window, producing
context_overflow0/5.
These are committed artifact-backed measurements. They do not establish a horizon multiplier; the harness needed to measure that target does not exist yet.
The sidecar keeps recent turns verbatim, compacts older spans into summaries plus a facts ledger, and assembles a fit-to-budget context for each request. Compaction runs in the background; a selector chooses which durable facts and episodes return to the model's desk.
The next harness separates three concerns:
- Brain: a fixed local model chooses actions.
- Notebook: naive, truncate, summary, and REM memory policies receive the same run.
- World: a deterministic environment emits observations and grades the task.
The canonical illustration is a long water-quality audit. The environment can be packaged with OpenEnv; an append-only event tape can replay the same run through every memory policy; and a state card can hand off goal, progress, key facts, and open items before the context budget is exhausted. See Lesson 4 §6.
New to this hardware? docs/hardware-context.md covers
what the NPU is, how it relates to the CPU and iGPU it shares the chip with, and
which of its behaviors you can actually observe — enough to evaluate REM's
placement claim or to build your own NPU use-case on the same foundation.
pip install -e ".[dev]"
pytestBenchmark entry points, artifact locations, and hardware prerequisites are documented in the placement benchmark and the implementation roadmap.
- Run the answerer/arm sweep on other hardware — and report the budget line. Include tokens injected per step, prompt-processing latency, and peak RAM, not only accuracy.
- Share tool-scaffolding prior art for reliable ~2B tool calls. Deterministic ordering, date, and abstention tools should clear the reasoning wall without spending the context budget on a larger answerer.
- Confirm OpenEnv fits fat-observation, ≥100-step tasks. We need evidence that its transport and environment model suit long document-heavy runs.
- Help design a cheap typed-identity judge. The bounded problem is deciding whether two facts describe the same slot without calling a model on every candidate pair.
- Share open-webui sidecar integration experience. Pipeline/filter-hook experience would shorten the path to daily dogfooding.
The rationale and contribution surface are in the direction review.
MIT License — see LICENSE. Free to use, modify, and distribute, including commercially; keep the copyright and permission notice. Author contact: title22.org.