What changes when agents remember.
Eight concrete scenarios where persistent memory changes how operators and agents work together. Each item lists the feature reference and the version it ships in.
-
Context that survives session boundaries
Every new session starts from zero. The agent has no memory of prior decisions, ongoing projects, or established working agreements.
Decisions, learned patterns, and project state persist in the vault and are surfaced automatically at the start of each relevant session.
-
Workflows that improve without manual tuning
Operator-level preferences and validated sequences stay implicit — repeated verbally, forgotten on model switch, never formalized.
Recurring patterns (dry-run before spec, validate before implementing) are distilled into SKILL.md entries and applied automatically.
-
Consistent agent identity across model switches
Each LLM backend reads a different config file. A gateway fallback silently changes the agent's tone, rules, or working style.
A single vault source of truth is served to every model. The agent's identity, charter, and hard limits are model-agnostic.
-
Shared rules that apply to operator and agent alike
Constraints live in prompts — overridden by context, silently dropped on reformat, inconsistently applied across sessions.
A structured constitution (atomic blocks selected at install) defines shared principles for operator, agent, and sub-agents. Only the operator can amend it. Only relevant blocks are injected per turn.
-
Your decisions outrank generic web content
A scraped article and an explicit architectural decision surface with equal weight. Generic content contradicts your own choices.
Explicit decisions carry trust 0.95, decay slowly (~2 years). Scraped content carries trust 0.35, near-invisible after 2 weeks. At equal semantic similarity, your decision always wins.
-
Confident-sounding wrong answers caught before you see them
A single aggregate score passes a threshold even when attribution is weak and specificity is high — a confident hallucination.
Attribution and specificity are scored independently. High specificity + low attribution triggers automatic REJECT + grounding retry. Scorer disagreement triggers review queue.
-
Relevant past decisions surfaced without asking
Operator re-decides what was already decided. Contradictory choices accumulate silently across sessions and projects.
When working on a problem, gradatum surfaces prior vault entries matching the current context. Contextual retrieval today; proactive push without explicit query planned for a later release.
-
Search that understands where a note lives, not just what it says
Standard vector RAG returns decontextualized chunks. A note about "write hooks" in an architecture section ranks equally with one in a debug log.
Each note is embedded with its structural path (e.g., "knowledge > architecture > writehook"). Section-aware retrieval surfaces notes from the right locus first.