diff --git a/pkg/scriptmem/memory.go b/pkg/scriptmem/memory.go
new file mode 100644
index 0000000..32373ac
--- /dev/null
+++ b/pkg/scriptmem/memory.go
@@ -0,0 +1,189 @@
+// Package script — memory.go is the bridge that attaches the existing
+// in-memory runtime to the unified pipeline. A `memory` block resolves
+// through the same Parse → Resolve front as everything else, then — instead
+// of lowering to a Sibyl Plan (the temporal path) — is handed to the
+// in-process interpreter that already implements every historical verb.
+//
+// Nothing about the runtime changes: it still consumes the old *Program
+// AST and runs exactly as it always has. This file only ADAPTS the new
+// resolved.AST into that *Program and calls the existing Execute. The two
+// backends now share one front end (translate/parse/resolve) and diverge
+// only at the final interpreter — temporal lowers+submits, memory adapts
+// +executes — which is the whole point of the unification.
+package scriptmem
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/vinodhalaharvi/agentscript/internal/agentscript"
+	"github.com/vinodhalaharvi/agentscript/pkg/script/ast"
+	"github.com/vinodhalaharvi/agentscript/pkg/script/resolved"
+)
+
+// MemoryConfig configures the in-memory runtime for a RunMemory call. It
+// mirrors the runtime's own config; all fields are optional (a verb that
+// needs a credential the config doesn't supply fails at execution, as it
+// always has).
+type MemoryConfig struct {
+	GeminiAPIKey       string
+	ClaudeAPIKey       string
+	SearchAPIKey       string
+	Model              string
+	Verbose            bool
+	GoogleCredsFile    string
+	GoogleTokenFile    string
+	GitHubClientID     string
+	GitHubClientSecret string
+	GitHubTokenFile    string
+}
+
+// RunMemory executes a resolved memory-backend program in-process and
+// returns its result. It is the synchronous sibling of the temporal
+// submit path: where temporal starts a durable workflow and returns a
+// handle, RunMemory runs the interpreter here and now and returns the
+// output directly.
+//
+// It expects the block to target the memory backend (callers route on the
+// backend keyword); a non-memory block is rejected so a temporal program
+// is never silently run in-process.
+func RunMemory(ctx context.Context, cfg MemoryConfig, r resolved.AST) (string, error) {
+	prog, backend, err := toProgram(r)
+	if err != nil {
+		return "", err
+	}
+	if backend != ast.BackendMemory {
+		return "", fmt.Errorf("RunMemory: program targets the %s backend, not memory", backend.String())
+	}
+
+	rt, err := agentscript.NewRuntime(ctx, agentscript.RuntimeConfig{
+		GeminiAPIKey:       cfg.GeminiAPIKey,
+		ClaudeAPIKey:       cfg.ClaudeAPIKey,
+		SearchAPIKey:       cfg.SearchAPIKey,
+		Model:              cfg.Model,
+		Verbose:            cfg.Verbose,
+		GoogleCredsFile:    cfg.GoogleCredsFile,
+		GoogleTokenFile:    cfg.GoogleTokenFile,
+		GitHubClientID:     cfg.GitHubClientID,
+		GitHubClientSecret: cfg.GitHubClientSecret,
+		GitHubTokenFile:    cfg.GitHubTokenFile,
+	})
+	if err != nil {
+		return "", fmt.Errorf("RunMemory: runtime init: %w", err)
+	}
+	return rt.Execute(ctx, prog)
+}
+
+// toProgram adapts a resolved.AST into the in-memory runtime's *Program.
+// The runtime is unchanged; this shim translates the unified pipeline's
+// representation into the one Execute already understands. It returns the
+// program and its backend so the caller can verify routing.
+//
+// The MVP pipeline produces exactly one block; multi-block programs are
+// rejected (the temporal path has the same single-block assumption).
+func toProgram(r resolved.AST) (*agentscript.Program, ast.Backend, error) {
+	if len(r.Blocks) != 1 {
+		return nil, 0, fmt.Errorf("toProgram: expected exactly one block, got %d", len(r.Blocks))
+	}
+	b := r.Blocks[0]
+	stmt, err := nodeToStatement(b.Body)
+	if err != nil {
+		return nil, 0, err
+	}
+	return &agentscript.Program{Statements: []*agentscript.Statement{stmt}}, b.Backend, nil
+}
+
+// nodeToStatement converts a resolved.Node into the old *Statement tree.
+//   - Pipeline → a chain of Statements linked by .Pipe (a >=> b >=> c)
+//   - Parallel → a Statement whose .Parallel holds the branches
+//   - Call     → a Statement whose .Command holds the verb + args
+func nodeToStatement(n resolved.Node) (*agentscript.Statement, error) {
+	switch node := n.(type) {
+	case resolved.Pipeline:
+		return pipelineToStatement(node)
+	case resolved.Parallel:
+		return parallelToStatement(node)
+	case resolved.Call:
+		cmd, err := callToCommand(node)
+		if err != nil {
+			return nil, err
+		}
+		return &agentscript.Statement{Command: cmd}, nil
+	default:
+		return nil, fmt.Errorf("nodeToStatement: unsupported node %T", n)
+	}
+}
+
+// pipelineToStatement folds [s0, s1, s2] into s0 >=> s1 >=> s2 by linking
+// each stage's .Pipe to the next. A single-stage pipeline is just that
+// stage.
+func pipelineToStatement(p resolved.Pipeline) (*agentscript.Statement, error) {
+	if len(p.Stages) == 0 {
+		return nil, fmt.Errorf("pipelineToStatement: empty pipeline")
+	}
+	stmts := make([]*agentscript.Statement, len(p.Stages))
+	for i, s := range p.Stages {
+		st, err := nodeToStatement(s)
+		if err != nil {
+			return nil, err
+		}
+		stmts[i] = st
+	}
+	// Link i → i+1 via Pipe. A stage that is itself a pipeline/parallel
+	// statement keeps its own structure; we attach the continuation to
+	// the tail of the chain.
+	for i := 0; i < len(stmts)-1; i++ {
+		tail := stmts[i]
+		for tail.Pipe != nil {
+			tail = tail.Pipe
+		}
+		tail.Pipe = stmts[i+1]
+	}
+	return stmts[0], nil
+}
+
+// parallelToStatement maps resolved.Parallel branches into the old
+// *Parallel form ( b0 <*> b1 <*> ... ).
+func parallelToStatement(p resolved.Parallel) (*agentscript.Statement, error) {
+	branches := make([]*agentscript.Statement, len(p.Branches))
+	for i, br := range p.Branches {
+		st, err := nodeToStatement(br)
+		if err != nil {
+			return nil, err
+		}
+		branches[i] = st
+	}
+	return &agentscript.Statement{Parallel: &agentscript.Parallel{Branches: branches}}, nil
+}
+
+// callToCommand maps a resolved.Call into the old *Command. The old form
+// holds up to four positional string args; the pipeline's permissive
+// memory specs are variadic-string, so we map the first four and reject
+// anything beyond what the old command shape can carry.
+func callToCommand(c resolved.Call) (*agentscript.Command, error) {
+	args := make([]string, 0, len(c.Args))
+	for _, a := range c.Args {
+		s, ok := a.(ast.StringArg)
+		if !ok {
+			return nil, fmt.Errorf("callToCommand: %s: non-string argument %T", c.Name, a)
+		}
+		args = append(args, s.Value)
+	}
+	if len(args) > 4 {
+		return nil, fmt.Errorf("callToCommand: %s: %d args exceeds the 4 the in-memory command form supports", c.Name, len(args))
+	}
+	cmd := &agentscript.Command{Action: c.Name}
+	if len(args) > 0 {
+		cmd.Arg = args[0]
+	}
+	if len(args) > 1 {
+		cmd.Arg2 = args[1]
+	}
+	if len(args) > 2 {
+		cmd.Arg3 = args[2]
+	}
+	if len(args) > 3 {
+		cmd.Arg4 = args[3]
+	}
+	return cmd, nil
+}
diff --git a/pkg/scriptmem/memory_adapter_test.go b/pkg/scriptmem/memory_adapter_test.go
new file mode 100644
index 0000000..9480eb8
--- /dev/null
+++ b/pkg/scriptmem/memory_adapter_test.go
@@ -0,0 +1,75 @@
+package scriptmem
+
+import (
+	"context"
+	"testing"
+
+	"github.com/vinodhalaharvi/agentscript/internal/agentscript"
+	"github.com/vinodhalaharvi/agentscript/pkg/script"
+	"github.com/vinodhalaharvi/agentscript/pkg/script/ast"
+)
+
+// Build a resolved memory program by running the real front end
+// (script.Parse → script.Resolve) against the complete registry, then
+// assert the adapter maps it to the correct old *Program shape. Exercises
+// the bridge's structural translation without needing a live runtime.
+func resolveMemory(t *testing.T, src string) (prog *agentscript.Program, backend ast.Backend) {
+	t.Helper()
+	a, err := script.Parse(context.Background(), script.Source(src))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	r, err := script.Resolve(context.Background(), script.CompleteRegistry(), a)
+	if err != nil {
+		t.Fatalf("resolve: %v", err)
+	}
+	prog, backend, err = toProgram(r)
+	if err != nil {
+		t.Fatalf("toProgram: %v", err)
+	}
+	return prog, backend
+}
+
+func TestAdapter_SingleCall(t *testing.T) {
+	prog, backend := resolveMemory(t, `memory static ( hf_summarize "the thread" )`)
+	if backend != ast.BackendMemory {
+		t.Fatalf("backend = %v, want memory", backend)
+	}
+	if len(prog.Statements) != 1 {
+		t.Fatalf("statements = %d, want 1", len(prog.Statements))
+	}
+	cmd := prog.Statements[0].Command
+	if cmd == nil {
+		t.Fatal("expected a Command")
+	}
+	if cmd.Action != "hf_summarize" {
+		t.Errorf("Action = %q, want hf_summarize", cmd.Action)
+	}
+	if cmd.Arg != "the thread" {
+		t.Errorf("Arg = %q, want 'the thread'", cmd.Arg)
+	}
+}
+
+func TestAdapter_PipelineChains(t *testing.T) {
+	prog, _ := resolveMemory(t, `memory static ( search "x" >=> hf_summarize >=> echo )`)
+	s := prog.Statements[0]
+	names := []string{}
+	for s != nil {
+		if s.Command != nil {
+			names = append(names, s.Command.Action)
+		}
+		s = s.Pipe
+	}
+	if len(names) != 3 || names[0] != "search" || names[1] != "hf_summarize" || names[2] != "echo" {
+		t.Errorf("pipeline chain = %v, want [search hf_summarize echo]", names)
+	}
+}
+
+func TestAdapter_RejectsTemporalBackend(t *testing.T) {
+	a, _ := script.Parse(context.Background(), script.Source(`temporal static ( echo "x" )`))
+	r, _ := script.Resolve(context.Background(), script.CompleteRegistry(), a)
+	_, err := RunMemory(context.Background(), MemoryConfig{}, r)
+	if err == nil {
+		t.Fatal("RunMemory should reject a temporal-backend program")
+	}
+}