Cyrene_v2/internal/runtime/loop/memory.go

package loop

import (
	"context"
	"strings"

	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
)

// injectMemory retrieves the persona's relevant long-term memories and prepends
// them as a leading system message, so the model can ground its answer in recalled
// context (design §6 / §2). It is re-run every round via assembleContext so the
// retrieval reflects the latest conversation rather than a first-turn snapshot. It
// returns the input messages unchanged when no retriever/persona is configured,
// retrieval fails, or nothing relevant is found.
func (l *Loop) injectMemory(ctx context.Context, messages []model.Message) []model.Message {
	if l.retriever == nil || l.personaID == "" {
		return messages
	}
	block, err := l.retriever.Inject(ctx, l.personaID, queryFromMessages(messages), 0)
	if err != nil {
		l.log.Warn("memory inject", "error", err)
		return messages
	}
	if block == "" {
		return messages
	}
	out := make([]model.Message, 0, len(messages)+1)
	out = append(out, model.TextMessage(model.RoleSystem, block))
	out = append(out, messages...)
	return out
}

// assembleContext builds the model-visible message list for one round (design
// §1.3 每轮 assemble, the loop's equivalent of OpenClaw's transformContext): it
// re-injects the persona's relevant long-term memory — retrieval is re-run every
// round from the current conversation, not cached from the first turn, so it
// supports cross-round refresh — folds in the per-session working-memory summary
// when present, and applies the optional TransformContext hook last so a caller
// can project/filter what the model actually sees. It returns the input unchanged
// when nothing is configured.
func (l *Loop) assembleContext(ctx context.Context, messages []model.Message, round int) []model.Message {
	var leading []model.Message
	if l.working != nil {
		if s := l.working.Summarize(); s != "" {
			leading = append(leading, model.TextMessage(model.RoleSystem, "工作记忆: "+s))
		}
	}
	leading = append(leading, l.injectMemory(ctx, messages)...)

	out := leading
	if l.transformContext != nil {
		out = l.transformContext(ctx, leading, round)
	}
	return out
}

// compactIfNeeded compacts the conversation when its estimated context exceeds
// the configured MaxContextTokens. It is a no-op when compaction is disabled
// (non-positive MaxContextTokens), the compactor is nil, the estimate is within
// budget, or the compactor returns an error — a compaction failure is logged and
// the conversation is kept as-is rather than failing the run.
func (l *Loop) compactIfNeeded(ctx context.Context, messages []model.Message) []model.Message {
	if l.maxContextTokens <= 0 || l.compactor == nil {
		return messages
	}
	compacted, err := l.compactor.Compact(ctx, messages, l.maxContextTokens)
	if err != nil {
		l.log.Warn("context compact", "error", err, "tokens", l.maxContextTokens)
		return messages
	}
	if compacted == nil {
		return messages
	}
	return compacted
}

// extractMemory runs the extraction hook (design §6) over the current
// conversation and persists any derived items. Errors are logged, never fatal, so
// a bad extraction cannot abort the run.
func (l *Loop) extractMemory(ctx context.Context, messages []model.Message, step int) {
	if l.extractor == nil || l.memory == nil {
		return
	}
	items, err := l.extractor.Extract(ctx, messages)
	if err != nil {
		l.log.Warn("memory extract", "step", step, "error", err)
		return
	}
	for _, it := range items {
		if it.PersonaID == "" && l.personaID != "" {
			it.PersonaID = l.personaID
		}
		if _, err := l.memory.Add(ctx, it); err != nil {
			l.log.Warn("memory store", "step", step, "error", err)
		}
	}
}

// queryFromMessages derives a retrieval query from the conversation: the last
// user turn's text, falling back to the last message's text, then "".
func queryFromMessages(messages []model.Message) string {
	for i := len(messages) - 1; i >= 0; i-- {
		if messages[i].Role == model.RoleUser {
			return messageText(messages[i])
		}
	}
	if n := len(messages); n > 0 {
		return messageText(messages[n-1])
	}
	return ""
}

// messageText returns the concatenated plain-text payload of a message's text
// parts, which is the shared view used to build a retrieval query.
func messageText(m model.Message) string {
	var b strings.Builder
	for _, part := range m.Content {
		if part.Type != model.ContentTypeText {
			continue
		}
		if b.Len() > 0 {
			b.WriteByte(' ')
		}
		b.WriteString(part.Text)
	}
	return b.String()
}