ACLA Agentic Context RuntimeRemember more. Process less.
AI agents do not only have a memory problem. They have a context-selection problem.
ACLA keeps persistent project context and compiles a Mission Working Set for each model call: the minimum relevant, current, authoritative and provenance-aware context that mission needs.
More stored context does not mean better reasoning.
Agents are increasingly given long-lived memory: project history, decisions, conversations and captured activity. Retaining that material is useful. Passing all of it, or a similarity-ranked slice of it, to the model is not the same as giving the model what it needs.
Stale
A superseded decision is still retrievable and still reads as true.
Conflicting
Two sources disagree and nothing decides which one the model should believe.
Weak authority
An informal remark carries the same weight as an approved decision.
Irrelevant
Material is accurate but does not bear on the question being asked.
Token waste
Budget is spent on context that cannot change the answer.
02 Memory ≠ context
An agent may remember far more than it should process for one mission.
Persistent context
What the project retains
Long-lived and growing
Holds current, superseded and historical records
May contain sources that disagree
Not bounded by a single model call
Mission Working Set
What the model receives for this mission
Compiled per mission
Minimum relevant and current
Weighted by authority, with conflicts handled explicitly
Provenance-aware and bounded by a token budget
ACLA keeps these two deliberately separate. Retaining more is cheap. Putting it in front of the model has a cost in tokens, attention and correctness.
03 Architecture
Context is compiled, not accumulated.
RetainPersistent Context
Project knowledge, decisions and captured workspace activity, including superseded and conflicting records.
DecideContext Compiler
Deterministic selection using:
relevance
authority
temporal state
provenance
conflict rules
mandatory context
token budget
CompileMission Working Set
minimum relevant
current
authoritative
provenance-aware
ReasonModel
Receives the working set, not the whole memory.
Adjacent, separate concerns
MCRcapability, provider and tool selection
EAFauthorization and outcome assurance
ACLA answers one question: what should the model know for this mission? Capability selection and authorization sit alongside it as separate concerns, not as ACLA components.
04 Context decisions
Seven questions the compiler answers before the model sees anything.
Relevance
Does this item bear on the mission question, or is it only nearby?
Authority
Who or what asserted it, and how much weight should that source carry?
Temporal state
Is it current, superseded or historical at the time of the mission?
Provenance
Where did it come from, and can an answer be traced back to it?
Conflicts
When sources disagree, which rule decides, and is the disagreement made visible rather than averaged away?
Mandatory context
What must always be present for this project, regardless of score?
Token constraints
What still fits the budget once mandatory context is placed?
Context sources can be registered with a temporal state and an authority class, so these decisions rest on declared properties rather than on similarity alone.
05 Runtime
A local runtime, not a library call.
ACLA runs as a set of local services with a CLI, a REST interface, a local UI and an MCP server.
Project context
Register a project against a workspace. Persistent context is kept per project and can be listed, added to or archived.
acla project add · acla context
Providers
Model providers are configured separately from projects. The release was qualified with a local model (qwen3:8b).
acla provider add
Missions
A mission is a question run against a project: the compiler builds the working set, then the model is called with it.
acla mission run
Lifecycle
Local services start, report independent component status, restart and stop. Stopping preserves durable state.
acla start · status · stop
Activity Bus
A local event bus for client activity, supervised by the runtime lifecycle.
acla activity-bus
Capture
Workspace capture observes a registered project read-only.
acla capture
MCP
ACLA tools are served to MCP-compatible clients over stdio with the [mcp] extra installed.
acla mcp serve
Replay and evidence
Mission runs are recorded so they can be inspected and replayed deterministically.
acla view-run · replay-run
06 Governance boundary
Three questions, kept apart.
ACLA
What should the agent know?
Selects and compiles mission context from persistent project context.
MCR
What capability, provider or tool should satisfy the mission?
Capability selection is a separate architectural concern.
EAF
What is the agent authorized to do, and how is the outcome proven?
Authorization and outcome assurance are a separate architectural concern.
Knowing is not the same as being able, and being able is not the same as being allowed. Collapsing these into one component makes each harder to reason about and harder to prove. Capella is a governed enterprise engineering environment where these ideas can be applied together.
07 Evidence
Release 4.0.0 qualification.
101 passed · 5 skippedSource suite and installed-wheel suite
7 / 7Lifecycle regression, source and installed
PassWindows start, restart and stop
2 missions + replayReal missions with deterministic replay and durable state
PassClean install from public PyPI, import isolation and CLI smoke
Scope: Windows with a local model (qwen3:8b). The Activity Bus was live and workspace capture was live in read-only mode. Published artefacts were read back byte for byte from the GitHub release and from PyPI, and the PyPI files were checked for hash and non-yanked status.
Linux and macOS lifecycle behaviour is not qualified for 4.0.0: one Linux lifecycle test did not pass because the model identity check remained pending. This is release evidence for a local runtime. It is not SAP, production, enterprise or provider certification, and it is not a claim of general performance improvement.
The example uses a local Ollama provider; install Ollama and pull qwen3:8b first. Replace /path/to/workspace with your project folder.
acla init
acla project add MYPROJECT --workspace /path/to/workspace
acla provider add local --type ollama --model qwen3:8b
acla start
acla status
acla mission run "What is the current blocker?" --project MYPROJECT
acla stop
This is the same path used to qualify the release: register a project, run a real mission against a local model, then inspect and replay the recorded run. A recorded walkthrough is not yet published.
Known issue in 4.0.0:acla status prints a stale release-status line that describes the build as an unreleased candidate. The published 4.0.0 artefacts are the ones listed under Evidence.
ACLA grew out of research into agentic context lifecycle architecture: contracts, compiler logic, evaluation and metrics for how context should move through an agent's work. Related published work treats context, token and tool use as engineering controls.
The runtime is an engineering implementation of those ideas. It does not re-run, revise or extend previously published research results; those stand as published.
